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> 5674ae819SStefano Zampini 6a3df083aSStefano Zampini #undef __FUNCT__ 7a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 8a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 9a3df083aSStefano Zampini { 10a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 11a3df083aSStefano Zampini PetscErrorCode ierr; 12a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 13a3df083aSStefano Zampini 14a3df083aSStefano Zampini PetscFunctionBegin; 15a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 16a3df083aSStefano Zampini if (transpose) { 17a3df083aSStefano Zampini apply_right = ctx->apply_left; 18a3df083aSStefano Zampini apply_left = ctx->apply_right; 19a3df083aSStefano Zampini } else { 20a3df083aSStefano Zampini apply_right = ctx->apply_right; 21a3df083aSStefano Zampini apply_left = ctx->apply_left; 22a3df083aSStefano Zampini } 23a3df083aSStefano Zampini reset_x = PETSC_FALSE; 24a3df083aSStefano Zampini if (apply_right) { 25a3df083aSStefano Zampini const PetscScalar *ax; 26a3df083aSStefano Zampini PetscInt nl,i; 27a3df083aSStefano Zampini 28a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 29a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 30a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 31a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 32a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 33a3df083aSStefano Zampini PetscScalar sum,val; 34a3df083aSStefano Zampini const PetscInt *idxs; 35a3df083aSStefano Zampini PetscInt nz,j; 36a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 37a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 38a3df083aSStefano Zampini sum = 0.; 39a3df083aSStefano Zampini if (ctx->apply_p0) { 40a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 41a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 42a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 43a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 44a3df083aSStefano Zampini } 45a3df083aSStefano Zampini } else { 46a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 47a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 48a3df083aSStefano Zampini } 49a3df083aSStefano Zampini } 50a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 51a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 52a3df083aSStefano Zampini } 53a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 54a3df083aSStefano Zampini reset_x = PETSC_TRUE; 55a3df083aSStefano Zampini } 56a3df083aSStefano Zampini if (transpose) { 57a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 58a3df083aSStefano Zampini } else { 59a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 60a3df083aSStefano Zampini } 61a3df083aSStefano Zampini if (reset_x) { 62a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 63a3df083aSStefano Zampini } 64a3df083aSStefano Zampini if (apply_left) { 65a3df083aSStefano Zampini PetscScalar *ay; 66a3df083aSStefano Zampini PetscInt i; 67a3df083aSStefano Zampini 68a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 69a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 70a3df083aSStefano Zampini PetscScalar sum,val; 71a3df083aSStefano Zampini const PetscInt *idxs; 72a3df083aSStefano Zampini PetscInt nz,j; 73a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 74a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 75a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 76a3df083aSStefano Zampini if (ctx->apply_p0) { 77a3df083aSStefano Zampini sum = 0.; 78a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 79a3df083aSStefano Zampini sum += ay[idxs[j]]; 80a3df083aSStefano Zampini ay[idxs[j]] += val; 81a3df083aSStefano Zampini } 82a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 83a3df083aSStefano Zampini } else { 84a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 85a3df083aSStefano Zampini ay[idxs[j]] += val; 86a3df083aSStefano Zampini } 87a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 88a3df083aSStefano Zampini } 89a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 90a3df083aSStefano Zampini } 91a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 92a3df083aSStefano Zampini } 93a3df083aSStefano Zampini PetscFunctionReturn(0); 94a3df083aSStefano Zampini } 95a3df083aSStefano Zampini 96a3df083aSStefano Zampini #undef __FUNCT__ 97a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 98a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 99a3df083aSStefano Zampini { 100a3df083aSStefano Zampini PetscErrorCode ierr; 101a3df083aSStefano Zampini 102a3df083aSStefano Zampini PetscFunctionBegin; 103a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 104a3df083aSStefano Zampini PetscFunctionReturn(0); 105a3df083aSStefano Zampini } 106a3df083aSStefano Zampini 107a3df083aSStefano Zampini #undef __FUNCT__ 108a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 109a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 110a3df083aSStefano Zampini { 111a3df083aSStefano Zampini PetscErrorCode ierr; 112a3df083aSStefano Zampini 113a3df083aSStefano Zampini PetscFunctionBegin; 114a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 115a3df083aSStefano Zampini PetscFunctionReturn(0); 116a3df083aSStefano Zampini } 117a3df083aSStefano Zampini 118a3df083aSStefano Zampini #undef __FUNCT__ 119a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 120a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 121a3df083aSStefano Zampini { 122a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 123a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 124a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 125a3df083aSStefano Zampini PetscErrorCode ierr; 126a3df083aSStefano Zampini 127a3df083aSStefano Zampini PetscFunctionBegin; 128a3df083aSStefano Zampini if (!restore) { 1291dd7afcfSStefano Zampini Mat A_IB,A_BI; 130a3df083aSStefano Zampini PetscScalar *work; 131df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 132a3df083aSStefano Zampini 1331dd7afcfSStefano Zampini if (pcbddc->benign_original_mat) { 1341dd7afcfSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 1351dd7afcfSStefano Zampini } 1361dd7afcfSStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) { 1371dd7afcfSStefano Zampini PetscFunctionReturn(0); 1381dd7afcfSStefano Zampini } 139a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 140a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 141a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 142a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 145a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 146a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 147a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 148a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 149a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 150a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 151059032f7SStefano Zampini if (reuse) { 152a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 1531dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 154059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 155059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 156059032f7SStefano Zampini PetscInt i; 157059032f7SStefano Zampini 158059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 159059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 160059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 161059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 162059032f7SStefano Zampini } 163059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 1641dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 165059032f7SStefano Zampini } 166a3df083aSStefano Zampini ctx->A = pcis->A_IB; 167a3df083aSStefano Zampini ctx->work = work; 168a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 169a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 170a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 171a3df083aSStefano Zampini pcis->A_IB = A_IB; 172a3df083aSStefano Zampini 173a3df083aSStefano Zampini /* A_BI as A_IB^T */ 174a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 175a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 176a3df083aSStefano Zampini pcis->A_BI = A_BI; 177a3df083aSStefano Zampini } else { 1781dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 1791dd7afcfSStefano Zampini PetscFunctionReturn(0); 1801dd7afcfSStefano Zampini } 181a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 182a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 183a3df083aSStefano Zampini pcis->A_IB = ctx->A; 1841dd7afcfSStefano Zampini ctx->A = NULL; 1851dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 1861dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 1871dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 1881dd7afcfSStefano Zampini if (ctx->free) { 189059032f7SStefano Zampini PetscInt i; 1901dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 191059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 192059032f7SStefano Zampini } 193059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 194059032f7SStefano Zampini } 195a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 196a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 197a3df083aSStefano Zampini } 198a3df083aSStefano Zampini PetscFunctionReturn(0); 199a3df083aSStefano Zampini } 200a3df083aSStefano Zampini 201a3df083aSStefano Zampini /* used just in bddc debug mode */ 202a3df083aSStefano Zampini #undef __FUNCT__ 203a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 204a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 205a3df083aSStefano Zampini { 206a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 207a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 208a3df083aSStefano Zampini Mat An; 209a3df083aSStefano Zampini PetscErrorCode ierr; 210a3df083aSStefano Zampini 211a3df083aSStefano Zampini PetscFunctionBegin; 212a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 213a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 214a3df083aSStefano Zampini if (is1) { 215a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 216a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 217a3df083aSStefano Zampini } else { 218a3df083aSStefano Zampini *B = An; 219a3df083aSStefano Zampini } 220a3df083aSStefano Zampini PetscFunctionReturn(0); 221a3df083aSStefano Zampini } 222a3df083aSStefano Zampini 2231cf9b237SStefano Zampini /* TODO: add reuse flag */ 2241cf9b237SStefano Zampini #undef __FUNCT__ 2251cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2261cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2271cf9b237SStefano Zampini { 2281cf9b237SStefano Zampini Mat Bt; 2291cf9b237SStefano Zampini PetscScalar *a,*bdata; 2301cf9b237SStefano Zampini const PetscInt *ii,*ij; 2311cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2321cf9b237SStefano Zampini PetscBool flg_row; 2331cf9b237SStefano Zampini PetscErrorCode ierr; 2341cf9b237SStefano Zampini 2351cf9b237SStefano Zampini PetscFunctionBegin; 2361cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2371cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2381cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2391cf9b237SStefano Zampini nnz = n; 2401cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2411cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2421cf9b237SStefano Zampini } 2431cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2441cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2451cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2461cf9b237SStefano Zampini nnz = 0; 2471cf9b237SStefano Zampini bii[0] = 0; 2481cf9b237SStefano Zampini for (i=0;i<n;i++) { 2491cf9b237SStefano Zampini PetscInt j; 2501cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2511cf9b237SStefano Zampini PetscScalar entry = a[j]; 2521cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2531cf9b237SStefano Zampini bij[nnz] = ij[j]; 2541cf9b237SStefano Zampini bdata[nnz] = entry; 2551cf9b237SStefano Zampini nnz++; 2561cf9b237SStefano Zampini } 2571cf9b237SStefano Zampini } 2581cf9b237SStefano Zampini bii[i+1] = nnz; 2591cf9b237SStefano Zampini } 2601cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2611cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2621cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2631cf9b237SStefano Zampini { 2641cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2651cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2661cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2671cf9b237SStefano Zampini } 2681cf9b237SStefano Zampini *B = Bt; 2691cf9b237SStefano Zampini PetscFunctionReturn(0); 2701cf9b237SStefano Zampini } 2711cf9b237SStefano Zampini 272674ae819SStefano Zampini #undef __FUNCT__ 2734f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2744f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2754f1b2e48SStefano Zampini { 2764f1b2e48SStefano Zampini Mat B; 2774f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2784f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2794f1b2e48SStefano Zampini PCBDDCGraph graph; 2804f1b2e48SStefano Zampini PetscInt i,n; 2814f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 2824f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 2834f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 2844f1b2e48SStefano Zampini PetscErrorCode ierr; 2854f1b2e48SStefano Zampini 2864f1b2e48SStefano Zampini PetscFunctionBegin; 28763c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 28863c961adSStefano Zampini *ncc = 0; 28963c961adSStefano Zampini *cc = NULL; 29063c961adSStefano Zampini PetscFunctionReturn(0); 29163c961adSStefano Zampini } 2924f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 2934f1b2e48SStefano Zampini if (!isseqaij && filter) { 2941cf9b237SStefano Zampini PetscBool isseqdense; 2951cf9b237SStefano Zampini 2961cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 2971cf9b237SStefano Zampini if (!isseqdense) { 2984f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 2991cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3001cf9b237SStefano Zampini PetscScalar *array; 3011cf9b237SStefano Zampini PetscReal chop=1.e-6; 3021cf9b237SStefano Zampini 3031cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3041cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3051cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3061cf9b237SStefano Zampini for (i=0;i<n;i++) { 3071cf9b237SStefano Zampini PetscInt j; 3081cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3091cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3101cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3111cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3121cf9b237SStefano Zampini } 3131cf9b237SStefano Zampini } 3141cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3151cf9b237SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); 3161cf9b237SStefano Zampini } 3174f1b2e48SStefano Zampini } else { 3184f1b2e48SStefano Zampini B = A; 3194f1b2e48SStefano Zampini } 3204f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3214f1b2e48SStefano Zampini 3224f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3234f1b2e48SStefano Zampini if (filter) { 3244f1b2e48SStefano Zampini PetscScalar *data; 3254f1b2e48SStefano Zampini PetscInt j,cum; 3264f1b2e48SStefano Zampini 3274f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3284f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3294f1b2e48SStefano Zampini cum = 0; 3304f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3314f1b2e48SStefano Zampini PetscInt t; 3324f1b2e48SStefano Zampini 3334f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3344f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3354f1b2e48SStefano Zampini continue; 3364f1b2e48SStefano Zampini } 3374f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini t = xadj_filtered[i]; 3404f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3414f1b2e48SStefano Zampini cum += t; 3424f1b2e48SStefano Zampini } 3434f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3444f1b2e48SStefano Zampini } else { 3454f1b2e48SStefano Zampini xadj_filtered = NULL; 3464f1b2e48SStefano Zampini adjncy_filtered = NULL; 3474f1b2e48SStefano Zampini } 3484f1b2e48SStefano Zampini 3494f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3504f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3514f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3524f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3534f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3544f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3554f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3564f1b2e48SStefano Zampini if (xadj_filtered) { 3574f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3584f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3594f1b2e48SStefano Zampini } else { 3604f1b2e48SStefano Zampini graph->xadj = xadj; 3614f1b2e48SStefano Zampini graph->adjncy = adjncy; 3624f1b2e48SStefano Zampini } 3634f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3644f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3654f1b2e48SStefano Zampini /* partial clean up */ 3664f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3674f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3681cf9b237SStefano Zampini if (A != B) { 3694f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3704f1b2e48SStefano Zampini } 3714f1b2e48SStefano Zampini 3724f1b2e48SStefano Zampini /* get back data */ 3731cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3741cf9b237SStefano Zampini if (cc) { 3754f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3764f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3774f1b2e48SStefano 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); 3784f1b2e48SStefano Zampini } 3794f1b2e48SStefano Zampini *cc = cc_n; 3801cf9b237SStefano Zampini } 3814f1b2e48SStefano Zampini /* clean up graph */ 3824f1b2e48SStefano Zampini graph->xadj = 0; 3834f1b2e48SStefano Zampini graph->adjncy = 0; 3844f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 3854f1b2e48SStefano Zampini PetscFunctionReturn(0); 3864f1b2e48SStefano Zampini } 3874f1b2e48SStefano Zampini 3884f1b2e48SStefano Zampini #undef __FUNCT__ 3895408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 3905408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 3915408967cSStefano Zampini { 3925408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3935408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 394dee84bffSStefano Zampini IS dirIS = NULL; 3954f1b2e48SStefano Zampini PetscInt i; 3965408967cSStefano Zampini PetscErrorCode ierr; 3975408967cSStefano Zampini 3985408967cSStefano Zampini PetscFunctionBegin; 399dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4005408967cSStefano Zampini if (zerodiag) { 4015408967cSStefano Zampini Mat A; 4025408967cSStefano Zampini Vec vec3_N; 4035408967cSStefano Zampini PetscScalar *vals; 4045408967cSStefano Zampini const PetscInt *idxs; 405d12d3064SStefano Zampini PetscInt nz,*count; 4065408967cSStefano Zampini 4075408967cSStefano Zampini /* p0 */ 4085408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4095408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4105408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4115408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4124f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4135408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4145408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4155408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4165408967cSStefano Zampini /* v_I */ 4175408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4185408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4195408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4205408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4215408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4225408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4235408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4245408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4255408967cSStefano Zampini if (dirIS) { 4265408967cSStefano Zampini PetscInt n; 4275408967cSStefano Zampini 4285408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4295408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4305408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4315408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4325408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4335408967cSStefano Zampini } 4345408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4355408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4365408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4375408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4385408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 4414f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 4425408967cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %f (should be numerically 0.)",PetscAbsScalar(vals[0])); 4435408967cSStefano Zampini } 4445408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4455408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 446d12d3064SStefano Zampini 447d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 448d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 449d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 450d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 451d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 452d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 453d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 454d12d3064SStefano Zampini if (count[idxs[i]]) { 455d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 456d12d3064SStefano Zampini } 457d4d8cf7bSStefano Zampini } 458d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 459d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4605408967cSStefano Zampini } 461dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4625408967cSStefano Zampini 4635408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4645408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4665408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4674f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4685408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4694f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4704f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4714f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 4724f1b2e48SStefano Zampini } 4735408967cSStefano Zampini } 4745408967cSStefano Zampini PetscFunctionReturn(0); 4755408967cSStefano Zampini } 4765408967cSStefano Zampini 4775408967cSStefano Zampini #undef __FUNCT__ 478339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 479339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 480339f8db1SStefano Zampini { 481339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4824f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 483b0f5fe93SStefano Zampini PetscInt nz,n; 4844f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 485339f8db1SStefano Zampini PetscErrorCode ierr; 486339f8db1SStefano Zampini 487339f8db1SStefano Zampini PetscFunctionBegin; 4889f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4899f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 490a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 491a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 492a3df083aSStefano Zampini } 493a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 494a3df083aSStefano Zampini pcbddc->benign_n = 0; 4954f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4964f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 4974f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 4984f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 4994f1b2e48SStefano Zampini since the local Schur complements are SPD 5004f1b2e48SStefano Zampini */ 5014f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5024f1b2e48SStefano Zampini have_null = PETSC_TRUE; 50340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5044f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5054f1b2e48SStefano Zampini 5064f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5074f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5084f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5094f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 510ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 51140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 51240fa8d13SStefano Zampini if (!sorted) { 51340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 51440fa8d13SStefano Zampini } 51540fa8d13SStefano Zampini } else { 51640fa8d13SStefano Zampini pressures = NULL; 51740fa8d13SStefano Zampini } 51897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 51997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 52097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 521339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 522339f8db1SStefano Zampini if (!sorted) { 523339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 524339f8db1SStefano Zampini } 525339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5264f1b2e48SStefano Zampini if (!nz) { 5274f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5284f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 52940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 53040fa8d13SStefano Zampini } 5314f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5324f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5334f1b2e48SStefano Zampini zerodiag_subs = NULL; 5344f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5354f1b2e48SStefano Zampini if (has_null_pressures) { 5364f1b2e48SStefano Zampini IS *subs; 5374f1b2e48SStefano Zampini PetscInt nsubs,i; 5384f1b2e48SStefano Zampini 5394f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5404f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5414f1b2e48SStefano Zampini if (nsubs > 1) { 5424f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5444f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5454f1b2e48SStefano Zampini IS t_zerodiag_subs; 5464f1b2e48SStefano Zampini PetscInt nl; 5474f1b2e48SStefano Zampini 5484f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5494f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5504f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5514f1b2e48SStefano Zampini if (nl) { 5524f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5534f1b2e48SStefano Zampini 5544f1b2e48SStefano Zampini if (pressures) { 5554f1b2e48SStefano Zampini IS t_pressure_subs; 5564f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5574f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5584f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5594f1b2e48SStefano Zampini } 5604f1b2e48SStefano Zampini if (valid) { 5614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini pcbddc->benign_n++; 5634f1b2e48SStefano Zampini } else { 5644f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5654f1b2e48SStefano Zampini } 5664f1b2e48SStefano Zampini } 5674f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini } 5704f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5714f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5724f1b2e48SStefano Zampini if (pressures) { 5734f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5744f1b2e48SStefano Zampini } 5754f1b2e48SStefano Zampini if (valid) { 5764f1b2e48SStefano Zampini pcbddc->benign_n = 1; 577ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5784f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5794f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5804f1b2e48SStefano Zampini } 5814f1b2e48SStefano Zampini } 5824f1b2e48SStefano Zampini } 5834f1b2e48SStefano Zampini 5844f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 5854f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5864f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5874f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5884f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5894f1b2e48SStefano Zampini } 5904f1b2e48SStefano Zampini 5914f1b2e48SStefano Zampini /* final check for null pressures */ 5924f1b2e48SStefano Zampini if (zerodiag && pressures) { 5934f1b2e48SStefano Zampini PetscInt nz,np; 5944f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5954f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 5964f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 5974f1b2e48SStefano Zampini } 5984f1b2e48SStefano Zampini 5994f1b2e48SStefano Zampini if (recompute_zerodiag) { 6004f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6014f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6024f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6044f1b2e48SStefano Zampini } else { 6054f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6064f1b2e48SStefano Zampini 6074f1b2e48SStefano Zampini nzn = 0; 6084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6094f1b2e48SStefano Zampini PetscInt ns; 6104f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6114f1b2e48SStefano Zampini nzn += ns; 6124f1b2e48SStefano Zampini } 6134f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6144f1b2e48SStefano Zampini nzn = 0; 6154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6164f1b2e48SStefano Zampini PetscInt ns,*idxs; 6174f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6184f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6194f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6204f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini nzn += ns; 6224f1b2e48SStefano Zampini } 6234f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6244f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6254f1b2e48SStefano Zampini } 6264f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6274f1b2e48SStefano Zampini } 6284f1b2e48SStefano Zampini 629b3afcdbeSStefano Zampini /* no-net-flux */ 630b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 631b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 632b3afcdbeSStefano Zampini MatNullSpace near_null_space; 633b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 634b3afcdbeSStefano Zampini Vec quad_vec; 635b3afcdbeSStefano Zampini PetscScalar *pvals; 636b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 637b3afcdbeSStefano Zampini 638b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 639b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 640b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 641b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 642b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 643b3afcdbeSStefano Zampini 644b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 645b3afcdbeSStefano Zampini np = 0; 646b3afcdbeSStefano Zampini if (zerodiag) { 647b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 648b3afcdbeSStefano Zampini } else if (pressures) { 649b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 650b3afcdbeSStefano Zampini } 651b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 652b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 653b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 654b3afcdbeSStefano Zampini if (np) { 655b3afcdbeSStefano Zampini const PetscInt *pidxs; 656b3afcdbeSStefano Zampini 657b3afcdbeSStefano Zampini if (zerodiag) { 658b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 659b3afcdbeSStefano Zampini } else if (pressures) { 660b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 661b3afcdbeSStefano Zampini } 662b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 663b3afcdbeSStefano Zampini if (zerodiag) { 664b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 665b3afcdbeSStefano Zampini } else if (pressures) { 666b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 667b3afcdbeSStefano Zampini } 668b3afcdbeSStefano Zampini } 669b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 670b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 671b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 672b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 673b3afcdbeSStefano Zampini 674b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 675b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 676b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = -1; 677b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 678b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 679b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 680b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 681b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 682b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 683b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) { 684b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 685b3afcdbeSStefano Zampini pvals[i] = 0.; 686b3afcdbeSStefano Zampini } 687b3afcdbeSStefano Zampini } 688b3afcdbeSStefano Zampini ierr = PetscMalloc1(matis->sf_nleaves,&dummyins);CHKERRQ(ierr); 689b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) dummyins[i] = i; 690b3afcdbeSStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf_nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 691b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 692b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 693b3afcdbeSStefano Zampini 694b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 695b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 696b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 697b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 698b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 699b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 700b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 701b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 702b3afcdbeSStefano Zampini } 703b3afcdbeSStefano Zampini 704b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7054f1b2e48SStefano Zampini if (has_null_pressures) { 7064f1b2e48SStefano Zampini IS zerodiagc; 7074f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7084f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7094f1b2e48SStefano Zampini 7104f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 711339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 712339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 713339f8db1SStefano Zampini /* local change of basis for pressures */ 714339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 71597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 716339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 717339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 718339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7194f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7204f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7214f1b2e48SStefano Zampini PetscInt nzs,j; 7224f1b2e48SStefano Zampini 7234f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7244f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7254f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7264f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7274f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7284f1b2e48SStefano Zampini } 729339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 730339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 731339f8db1SStefano Zampini /* set identity on velocities */ 732339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 733339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 734339f8db1SStefano Zampini } 7354f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7364f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7379f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7384f1b2e48SStefano 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); 739339f8db1SStefano Zampini /* set change on pressures */ 7404f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7414f1b2e48SStefano Zampini PetscScalar *array; 7424f1b2e48SStefano Zampini PetscInt nzs; 7434f1b2e48SStefano Zampini 7444f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7454f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7464f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 747339f8db1SStefano Zampini PetscScalar vals[2]; 748339f8db1SStefano Zampini PetscInt cols[2]; 749339f8db1SStefano Zampini 750339f8db1SStefano Zampini cols[0] = idxs[i]; 7514f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 752339f8db1SStefano Zampini vals[0] = 1.; 753b0f5fe93SStefano Zampini vals[1] = 1.; 7544f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 755339f8db1SStefano Zampini } 7564f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7574f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7584f1b2e48SStefano Zampini array[nzs-1] = 1.; 7594f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7604f1b2e48SStefano Zampini /* store local idxs for p0 */ 7614f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7624f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 763339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7644f1b2e48SStefano Zampini } 765339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 766339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 767a3df083aSStefano Zampini /* project if needed */ 768a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7691dd7afcfSStefano Zampini Mat M; 7701dd7afcfSStefano Zampini 7711dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 772339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 7731dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 7741dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 775a3df083aSStefano Zampini } 7764f1b2e48SStefano Zampini /* store global idxs for p0 */ 7774f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 778339f8db1SStefano Zampini } 779ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7804f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 781b0f5fe93SStefano Zampini 782b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 783b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 784339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 785339f8db1SStefano Zampini PetscFunctionReturn(0); 786339f8db1SStefano Zampini } 787339f8db1SStefano Zampini 788339f8db1SStefano Zampini #undef __FUNCT__ 789015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 790015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 791efc2fbd9SStefano Zampini { 792efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 793efc2fbd9SStefano Zampini PetscErrorCode ierr; 794efc2fbd9SStefano Zampini 795efc2fbd9SStefano Zampini PetscFunctionBegin; 796efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 797efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 7984f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 799efc2fbd9SStefano Zampini } 800015636ebSStefano Zampini if (get) { /* use SF to get values */ 801efc2fbd9SStefano Zampini PetscScalar *array; 802efc2fbd9SStefano Zampini 803efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8044f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8054f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 806efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8074f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 8084f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 809efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 810efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 811efc2fbd9SStefano Zampini } 812efc2fbd9SStefano Zampini PetscFunctionReturn(0); 813efc2fbd9SStefano Zampini } 814efc2fbd9SStefano Zampini 815efc2fbd9SStefano Zampini #undef __FUNCT__ 816c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 817c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 818c263805aSStefano Zampini { 819c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 820c263805aSStefano Zampini PetscErrorCode ierr; 821c263805aSStefano Zampini 822c263805aSStefano Zampini PetscFunctionBegin; 823c263805aSStefano Zampini /* TODO: add error checking 824c263805aSStefano Zampini - avoid nested pop (or push) calls. 825c263805aSStefano Zampini - cannot push before pop. 8261c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 827c263805aSStefano Zampini */ 8284f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 829efc2fbd9SStefano Zampini PetscFunctionReturn(0); 830efc2fbd9SStefano Zampini } 831c263805aSStefano Zampini if (pop) { 832a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8334f1b2e48SStefano Zampini IS is_p0; 8344f1b2e48SStefano Zampini MatReuse reuse; 835c263805aSStefano Zampini 836c263805aSStefano Zampini /* extract B_0 */ 8374f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8384f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8394f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8404f1b2e48SStefano Zampini } 8414f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8424f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 843c263805aSStefano Zampini /* remove rows and cols from local problem */ 844c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 84597d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8464f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8474f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 848a3df083aSStefano Zampini } else { 849a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 850a3df083aSStefano Zampini PetscScalar *vals; 851a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 852a3df083aSStefano Zampini 853a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 854a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 855a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8560b5adadeSStefano Zampini PetscInt *nnz; 857a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 858a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 859a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 860331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 861331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 862331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 863331e053bSStefano Zampini nnz[i] = n - nnz[i]; 864331e053bSStefano Zampini } 865331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 866331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 867331e053bSStefano Zampini } 868a3df083aSStefano Zampini 869a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 870a3df083aSStefano Zampini PetscScalar *array; 871a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 872a3df083aSStefano Zampini 873a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 874a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 875a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 876a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 877a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 878a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 879a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 880a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 881a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 882a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 883a3df083aSStefano Zampini cum = 0; 884a3df083aSStefano Zampini for (j=0;j<n;j++) { 885a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 886a3df083aSStefano Zampini vals[cum] = array[j]; 887a3df083aSStefano Zampini idxs_ins[cum] = j; 888a3df083aSStefano Zampini cum++; 889a3df083aSStefano Zampini } 890a3df083aSStefano Zampini } 891a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 892a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 893a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 894a3df083aSStefano Zampini } 895a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 896a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 897a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 898a3df083aSStefano Zampini } 899c263805aSStefano Zampini } else { /* push */ 900a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9014f1b2e48SStefano Zampini PetscInt i; 9024f1b2e48SStefano Zampini 9034f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9044f1b2e48SStefano Zampini PetscScalar *B0_vals; 9054f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9064f1b2e48SStefano Zampini 9074f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9084f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9097b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9104f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9114f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9124f1b2e48SStefano Zampini } 913c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 914c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 915a3df083aSStefano Zampini } else { 916a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 917a3df083aSStefano Zampini } 918c263805aSStefano Zampini } 919c263805aSStefano Zampini PetscFunctionReturn(0); 920c263805aSStefano Zampini } 921c263805aSStefano Zampini 922c263805aSStefano Zampini #undef __FUNCT__ 923b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 92408122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 925b1b3d7a2SStefano Zampini { 926b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 92708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 92808122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 92908122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 93008122e43SStefano Zampini PetscScalar *work,lwork; 93108122e43SStefano Zampini PetscScalar *St,*S,*eigv; 93208122e43SStefano Zampini PetscScalar *Sarray,*Starray; 93308122e43SStefano Zampini PetscReal *eigs,thresh; 9341b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 935f6f667cfSStefano Zampini PetscBool allocated_S_St; 93608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 93708122e43SStefano Zampini PetscReal *rwork; 93808122e43SStefano Zampini #endif 939b1b3d7a2SStefano Zampini PetscErrorCode ierr; 940b1b3d7a2SStefano Zampini 941b1b3d7a2SStefano Zampini PetscFunctionBegin; 942*af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 943*af25d912SStefano 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); 94406a4e24aSStefano Zampini 945fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 946fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 947fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 948fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9491575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 950fd14bc51SStefano Zampini } 951fd14bc51SStefano Zampini 952e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 953e496cd5dSStefano 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); 954e496cd5dSStefano Zampini } 955e496cd5dSStefano Zampini 95608122e43SStefano Zampini /* max size of subsets */ 95708122e43SStefano Zampini mss = 0; 95808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 95908122e43SStefano Zampini PetscInt subset_size; 960862806e4SStefano Zampini 96108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 96208122e43SStefano Zampini mss = PetscMax(mss,subset_size); 96308122e43SStefano Zampini } 96408122e43SStefano Zampini 96508122e43SStefano Zampini /* min/max and threshold */ 96608122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 967f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 96808122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 969f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 970f6f667cfSStefano Zampini if (nmin) { 971f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 972f6f667cfSStefano Zampini } 97308122e43SStefano Zampini 97408122e43SStefano Zampini /* allocate lapack workspace */ 97508122e43SStefano Zampini cum = cum2 = 0; 97608122e43SStefano Zampini maxneigs = 0; 97708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 97808122e43SStefano Zampini PetscInt n,subset_size; 979f6f667cfSStefano Zampini 98008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 98108122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9829162d606SStefano Zampini cum += subset_size; 9839162d606SStefano Zampini cum2 += subset_size*n; 98408122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 98508122e43SStefano Zampini } 98608122e43SStefano Zampini if (mss) { 9879ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 98808122e43SStefano Zampini PetscBLASInt B_itype = 1; 98908122e43SStefano Zampini PetscBLASInt B_N = mss; 9904c6709b3SStefano Zampini PetscReal zero = 0.0; 9914c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 99208122e43SStefano Zampini 99308122e43SStefano Zampini B_lwork = -1; 99408122e43SStefano Zampini S = NULL; 99508122e43SStefano Zampini St = NULL; 996a58a30b4SStefano Zampini eigs = NULL; 997a58a30b4SStefano Zampini eigv = NULL; 998a58a30b4SStefano Zampini B_iwork = NULL; 999a58a30b4SStefano Zampini B_ifail = NULL; 1000d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1001d1710679SStefano Zampini rwork = NULL; 1002d1710679SStefano Zampini #endif 10038bec7fa6SStefano Zampini thresh = 1.0; 100408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 100508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 100608122e43SStefano 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)); 100708122e43SStefano Zampini #else 100808122e43SStefano 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)); 100908122e43SStefano Zampini #endif 101008122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 101108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 101208122e43SStefano Zampini } else { 101308122e43SStefano Zampini /* TODO */ 101408122e43SStefano Zampini } 101508122e43SStefano Zampini } else { 101608122e43SStefano Zampini lwork = 0; 101708122e43SStefano Zampini } 101808122e43SStefano Zampini 101908122e43SStefano Zampini nv = 0; 1020d62866d3SStefano 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) */ 1021d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 102208122e43SStefano Zampini } 10234c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1024f6f667cfSStefano Zampini if (allocated_S_St) { 1025f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1026f6f667cfSStefano Zampini } 1027f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 102808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 102908122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 103008122e43SStefano Zampini #endif 10319162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10329162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10339162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 103408122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10359162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 103608122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 103708122e43SStefano Zampini 103808122e43SStefano Zampini maxneigs = 0; 103972b8c272SStefano Zampini cum = cumarray = 0; 10409162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10419162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1042d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 104308122e43SStefano Zampini const PetscInt *idxs; 104408122e43SStefano Zampini 1045d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 104608122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 104708122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 104808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 104908122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10509162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10519162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 105208122e43SStefano Zampini } 1053d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 105408122e43SStefano Zampini } 105508122e43SStefano Zampini 105608122e43SStefano Zampini if (mss) { /* multilevel */ 105708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 105808122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 105908122e43SStefano Zampini } 106008122e43SStefano Zampini 1061ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 106208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 106308122e43SStefano Zampini const PetscInt *idxs; 1064f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1065862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 106608122e43SStefano Zampini PetscBLASInt B_N; 1067aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 106808122e43SStefano Zampini 1069862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1070ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1071f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1072f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10739ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1074aff50787SStefano Zampini PetscInt j,k; 1075aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1076aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1077aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 107808122e43SStefano Zampini } 107908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1080aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1081aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1082aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1083aff50787SStefano Zampini } 108408122e43SStefano Zampini } 108508122e43SStefano Zampini } else { 108608122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 108708122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 108808122e43SStefano Zampini } 10898bec7fa6SStefano Zampini } else { 1090f6f667cfSStefano Zampini S = Sarray + cumarray; 1091f6f667cfSStefano Zampini St = Starray + cumarray; 10928bec7fa6SStefano Zampini } 1093aff50787SStefano Zampini /* see if we can save some work */ 1094aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1095aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1096aff50787SStefano Zampini } 1097aff50787SStefano Zampini 1098aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1099aff50787SStefano Zampini B_neigs = 0; 1100aff50787SStefano Zampini } else { 11019ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 110208122e43SStefano Zampini PetscBLASInt B_itype = 1; 1103f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11044c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11059552c7c7SStefano Zampini PetscInt nmin_s; 110608122e43SStefano Zampini 1107fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11088bec7fa6SStefano 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]]); 1109fd14bc51SStefano Zampini } 1110d16cbb6bSStefano Zampini 111108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1112d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1113d16cbb6bSStefano Zampini 1114d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 111508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1116f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 111708122e43SStefano Zampini #else 1118f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 111908122e43SStefano Zampini #endif 1120d16cbb6bSStefano Zampini } else { 1121d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1122d16cbb6bSStefano Zampini B_IL = 1; 1123d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1124d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1125d16cbb6bSStefano Zampini #else 1126d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1127d16cbb6bSStefano Zampini #endif 1128d16cbb6bSStefano Zampini } 112908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 113008122e43SStefano Zampini if (B_ierr) { 11316c4ed002SBarry 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); 11326c4ed002SBarry 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); 11336c4ed002SBarry 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); 113408122e43SStefano Zampini } 113508122e43SStefano Zampini 113608122e43SStefano Zampini if (B_neigs > nmax) { 1137fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1138fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1139fd14bc51SStefano Zampini } 1140f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 114108122e43SStefano Zampini B_neigs = nmax; 114208122e43SStefano Zampini } 114308122e43SStefano Zampini 11449552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11459552c7c7SStefano Zampini if (B_neigs < nmin_s) { 114608122e43SStefano Zampini PetscBLASInt B_neigs2; 114708122e43SStefano Zampini 1148f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1149f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1150fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1151fd14bc51SStefano 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); 1152fd14bc51SStefano Zampini } 11539ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 115408122e43SStefano Zampini PetscInt j; 115508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 115608122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 115708122e43SStefano Zampini } 115808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 115908122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 116008122e43SStefano Zampini } 116108122e43SStefano Zampini } else { 116208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 116308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 116408122e43SStefano Zampini } 116508122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 116608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1167f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&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)); 116808122e43SStefano Zampini #else 1169f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&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)); 117008122e43SStefano Zampini #endif 117108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 117208122e43SStefano Zampini B_neigs += B_neigs2; 117308122e43SStefano Zampini } 117408122e43SStefano Zampini if (B_ierr) { 11756c4ed002SBarry 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); 11766c4ed002SBarry 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); 11776c4ed002SBarry 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); 117808122e43SStefano Zampini } 1179fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1180ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 118108122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 118208122e43SStefano Zampini if (eigs[j] == 0.0) { 1183ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 118408122e43SStefano Zampini } else { 1185ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1186fd14bc51SStefano Zampini } 118708122e43SStefano Zampini } 118808122e43SStefano Zampini } 118908122e43SStefano Zampini } else { 119008122e43SStefano Zampini /* TODO */ 119108122e43SStefano Zampini } 1192aff50787SStefano Zampini } 11936c3e6151SStefano Zampini /* change the basis back to the original one */ 11946c3e6151SStefano Zampini if (sub_schurs->change) { 119572b8c272SStefano Zampini Mat change,phi,phit; 11966c3e6151SStefano Zampini 11976c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 11986c3e6151SStefano Zampini PetscInt ii; 11996c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12006c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12016c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12026c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12036c3e6151SStefano Zampini } 12046c3e6151SStefano Zampini } 12056c3e6151SStefano Zampini } 120672b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12076c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 120872b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12096c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12106c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12116c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12126c3e6151SStefano Zampini } 12138bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12148bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12159162d606SStefano Zampini if (B_neigs) { 12169162d606SStefano 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); 1217fd14bc51SStefano Zampini 1218fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12199552c7c7SStefano Zampini PetscInt ii; 12209552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1221ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12229552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1223ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1224ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1225ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1226ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1227ac47001eSStefano Zampini #else 1228ac47001eSStefano 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); 1229ac47001eSStefano Zampini #endif 12309552c7c7SStefano Zampini } 12319552c7c7SStefano Zampini } 1232fd14bc51SStefano Zampini } 12339162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12349162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12359162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12369162d606SStefano Zampini cum++; 123708122e43SStefano Zampini } 123808122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 123908122e43SStefano Zampini /* shift for next computation */ 124008122e43SStefano Zampini cumarray += subset_size*subset_size; 124108122e43SStefano Zampini } 1242fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1243fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1244fd14bc51SStefano Zampini } 124508122e43SStefano Zampini 124608122e43SStefano Zampini if (mss) { 124708122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 124808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1249f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1250f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1251f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 125208122e43SStefano Zampini } 1253f6f667cfSStefano Zampini if (allocated_S_St) { 1254f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1255f6f667cfSStefano Zampini } 1256f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 125708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 125808122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 125908122e43SStefano Zampini #endif 126008122e43SStefano Zampini if (pcbddc->dbg_flag) { 12611b968477SStefano Zampini PetscInt maxneigs_r; 1262b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12639b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 126408122e43SStefano Zampini } 126508122e43SStefano Zampini PetscFunctionReturn(0); 126608122e43SStefano Zampini } 1267b1b3d7a2SStefano Zampini 1268674ae819SStefano Zampini #undef __FUNCT__ 1269c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1270c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1271c8587f34SStefano Zampini { 1272c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12738629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1274c8587f34SStefano Zampini PetscErrorCode ierr; 1275c8587f34SStefano Zampini 1276c8587f34SStefano Zampini PetscFunctionBegin; 1277f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12785e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1279c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1280c8587f34SStefano Zampini 1281684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12820fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1283684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1284c8587f34SStefano Zampini 1285c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1286b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1287c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1288c8587f34SStefano Zampini } 1289c8587f34SStefano Zampini 12908629588bSStefano Zampini /* 12918629588bSStefano Zampini Setup local correction and local part of coarse basis. 12928629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12938629588bSStefano Zampini */ 129447f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12958629588bSStefano Zampini 12968629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12978629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12988629588bSStefano Zampini 12998629588bSStefano Zampini /* free */ 13008629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1301c8587f34SStefano Zampini PetscFunctionReturn(0); 1302c8587f34SStefano Zampini } 1303c8587f34SStefano Zampini 1304c8587f34SStefano Zampini #undef __FUNCT__ 1305674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1306674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1307674ae819SStefano Zampini { 1308674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1309674ae819SStefano Zampini PetscErrorCode ierr; 1310674ae819SStefano Zampini 1311674ae819SStefano Zampini PetscFunctionBegin; 1312674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1313674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 131430368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1315674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1316674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1317785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1318674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1319f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1320f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1321785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 132263602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 132363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1324674ae819SStefano Zampini PetscFunctionReturn(0); 1325674ae819SStefano Zampini } 1326674ae819SStefano Zampini 1327674ae819SStefano Zampini #undef __FUNCT__ 1328674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1329674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1330674ae819SStefano Zampini { 1331674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13324f1b2e48SStefano Zampini PetscInt i; 1333674ae819SStefano Zampini PetscErrorCode ierr; 1334674ae819SStefano Zampini 1335674ae819SStefano Zampini PetscFunctionBegin; 1336b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1337674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 133816909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 13391dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1340674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1341674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13424f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13434f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13444f1b2e48SStefano Zampini } 13454f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1346b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1347674ae819SStefano Zampini PetscFunctionReturn(0); 1348674ae819SStefano Zampini } 1349674ae819SStefano Zampini 1350674ae819SStefano Zampini #undef __FUNCT__ 1351674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1352674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1353674ae819SStefano Zampini { 1354674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1355674ae819SStefano Zampini PetscErrorCode ierr; 1356674ae819SStefano Zampini 1357674ae819SStefano Zampini PetscFunctionBegin; 1358674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 135958da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1360ca92afb2SStefano Zampini PetscScalar *array; 136106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 136206656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 136358da7f69SStefano Zampini } 1364674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1365674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 136615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 136715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1368674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1369674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1370674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 137106656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1372674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1373674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13748ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1375674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1376674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1377674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1378f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1379f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1380f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1381f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1382727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13830e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1384f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 138570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 13866e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 138781d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 13880369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 13891dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 13904f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13918b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1392ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1393ca92afb2SStefano Zampini PetscInt i; 1394ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1395ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1396ca92afb2SStefano Zampini } 1397ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1398ca92afb2SStefano Zampini } 13994f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1400674ae819SStefano Zampini PetscFunctionReturn(0); 1401674ae819SStefano Zampini } 1402674ae819SStefano Zampini 1403674ae819SStefano Zampini #undef __FUNCT__ 1404f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1405f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14066bfb1811SStefano Zampini { 14076bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14086bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14096bfb1811SStefano Zampini VecType impVecType; 14104f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14116bfb1811SStefano Zampini PetscErrorCode ierr; 14126bfb1811SStefano Zampini 14136bfb1811SStefano Zampini PetscFunctionBegin; 14146c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1415e7b262bdSStefano Zampini /* get sizes */ 14164f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1417b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14186bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1419e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1420e7b262bdSStefano Zampini /* R nodes */ 1421e7b262bdSStefano Zampini old_size = -1; 1422e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1423e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1424e7b262bdSStefano Zampini } 1425e7b262bdSStefano Zampini if (n_R != old_size) { 1426e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1427e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14286bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14296bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14306bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14316bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1432e7b262bdSStefano Zampini } 1433e7b262bdSStefano Zampini /* local primal dofs */ 1434e7b262bdSStefano Zampini old_size = -1; 1435e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1436e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1437e7b262bdSStefano Zampini } 1438e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1439e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 144083b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1441e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14426bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1443e7b262bdSStefano Zampini } 1444e7b262bdSStefano Zampini /* local explicit constraints */ 1445e7b262bdSStefano Zampini old_size = -1; 1446e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1447e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1448e7b262bdSStefano Zampini } 1449e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1450e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 145183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 145283b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 145383b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 145483b7ccabSStefano Zampini } 14556bfb1811SStefano Zampini PetscFunctionReturn(0); 14566bfb1811SStefano Zampini } 14576bfb1811SStefano Zampini 14586bfb1811SStefano Zampini #undef __FUNCT__ 145947f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 146047f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 146188ebb749SStefano Zampini { 146225084f0cSStefano Zampini PetscErrorCode ierr; 146325084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 146488ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 146588ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1466d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 146725084f0cSStefano Zampini /* submatrices of local problem */ 146880677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 146906656605SStefano Zampini /* submatrices of local coarse problem */ 147006656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 147125084f0cSStefano Zampini /* working matrices */ 147206656605SStefano Zampini Mat C_CR; 147325084f0cSStefano Zampini /* additional working stuff */ 147406656605SStefano Zampini PC pc_R; 14754f1b2e48SStefano Zampini Mat F; 14765cbda25cSStefano Zampini Vec dummy_vec; 1477a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 147825084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 147906656605SStefano Zampini PetscScalar *work; 148006656605SStefano Zampini PetscInt *idx_V_B; 1481ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 148206656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1483ffd830a3SStefano Zampini 148425084f0cSStefano Zampini /* some shortcuts to scalars */ 148506656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 148688ebb749SStefano Zampini 148788ebb749SStefano Zampini PetscFunctionBegin; 1488ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1489ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1490ffd830a3SStefano Zampini } 1491ffd830a3SStefano Zampini 1492ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1493b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14944f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1495b371cd4fSStefano Zampini n_B = pcis->n_B; 1496b371cd4fSStefano Zampini n_D = pcis->n - n_B; 149788ebb749SStefano Zampini n_R = pcis->n - n_vertices; 149888ebb749SStefano Zampini 149988ebb749SStefano Zampini /* vertices in boundary numbering */ 1500785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15010e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 15026c4ed002SBarry 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); 150388ebb749SStefano Zampini 150406656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1505019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 150606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 150706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 150806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 150906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 151006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 151106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 151206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 151306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 151406656605SStefano Zampini 151506656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 151606656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 151706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 151806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 151906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1520ffd830a3SStefano Zampini lda_rhs = n_R; 1521a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 152206656605SStefano Zampini if (isLU || isILU || isCHOL) { 152306656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1524df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1525df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1526d62866d3SStefano Zampini MatFactorType type; 1527d62866d3SStefano Zampini 1528df4d28bfSStefano Zampini F = reuse_solver->F; 15296816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1530d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1531ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1532df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 153306656605SStefano Zampini } else { 153406656605SStefano Zampini F = NULL; 153506656605SStefano Zampini } 153606656605SStefano Zampini 1537ffd830a3SStefano Zampini /* allocate workspace */ 1538ffd830a3SStefano Zampini n = 0; 1539ffd830a3SStefano Zampini if (n_constraints) { 1540ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1541ffd830a3SStefano Zampini } 1542ffd830a3SStefano Zampini if (n_vertices) { 1543ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1544ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1545ffd830a3SStefano Zampini } 1546ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1547ffd830a3SStefano Zampini 15485cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 15495cbda25cSStefano Zampini dummy_vec = NULL; 15505cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 15515cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 15525cbda25cSStefano Zampini } 15535cbda25cSStefano Zampini 155488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 155588ebb749SStefano Zampini if (n_constraints) { 155672b8c272SStefano Zampini Mat M1,M2,M3,C_B; 155706656605SStefano Zampini IS is_aux; 155880677318SStefano Zampini PetscScalar *array,*array2; 155906656605SStefano Zampini 1560f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 156180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 156288ebb749SStefano Zampini 156325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 156425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15658ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 156672b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 156788ebb749SStefano Zampini 156880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 156980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1570ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 157188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 157206656605SStefano Zampini const PetscScalar *row_cmat_values; 157306656605SStefano Zampini const PetscInt *row_cmat_indices; 157406656605SStefano Zampini PetscInt size_of_constraint,j; 157588ebb749SStefano Zampini 157606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 157706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1578ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 157906656605SStefano Zampini } 158006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 158106656605SStefano Zampini } 1582ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 158306656605SStefano Zampini if (F) { 158406656605SStefano Zampini Mat B; 158506656605SStefano Zampini 1586ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1587a3df083aSStefano Zampini if (need_benign_correction) { 1588df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1589a3df083aSStefano Zampini 159072b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 159172b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1592a3df083aSStefano Zampini } 159380677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1594a3df083aSStefano Zampini if (need_benign_correction) { 1595a3df083aSStefano Zampini PetscScalar *marr; 1596df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1597a3df083aSStefano Zampini 1598a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 15995cbda25cSStefano Zampini if (lda_rhs != n_R) { 16005cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 16015cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 16025cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 16035cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 16045cbda25cSStefano Zampini } 16055cbda25cSStefano Zampini } else { 1606a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1607a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 16085cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1609a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1610a3df083aSStefano Zampini } 16115cbda25cSStefano Zampini } 1612a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1613a3df083aSStefano Zampini } 161406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 161506656605SStefano Zampini } else { 161680677318SStefano Zampini PetscScalar *marr; 161780677318SStefano Zampini 161880677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 161906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1620ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1621ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 162206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 162306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 162406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 162506656605SStefano Zampini } 162680677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 162706656605SStefano Zampini } 162880677318SStefano Zampini if (!pcbddc->switch_static) { 162980677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 163080677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 163180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 163280677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1633ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 163480677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 163580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 163680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 163780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 163880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 163980677318SStefano Zampini } 164080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 164180677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 164272b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 164380677318SStefano Zampini } else { 1644ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1645ffd830a3SStefano Zampini IS dummy; 1646ffd830a3SStefano Zampini 1647ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 164872b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1649ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1650ffd830a3SStefano Zampini } else { 165180677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 165280677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1653ffd830a3SStefano Zampini } 165425084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 165580677318SStefano Zampini } 165680677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 165780677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 165880677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 165906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 166006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 166180677318SStefano Zampini if (isCHOL) { 166280677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 166380677318SStefano Zampini } else { 166425084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 166580677318SStefano Zampini } 166680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 166706656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 166825084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 166925084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 167025084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 167180677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 167272b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 167372b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 167406656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 167506656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1676f4ddd8eeSStefano Zampini } 1677fc227af8SStefano Zampini 1678fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 167988ebb749SStefano Zampini if (n_vertices) { 168006656605SStefano Zampini IS is_aux; 16813a50541eSStefano Zampini 1682df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16836816873aSStefano Zampini IS tis; 16846816873aSStefano Zampini 16856816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16866816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16876816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16886816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16896816873aSStefano Zampini } else { 16903a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16916816873aSStefano Zampini } 16929577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16939577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 169404708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 169525084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 169688ebb749SStefano Zampini } 169788ebb749SStefano Zampini 169888ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1699f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 170006656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 170106656605SStefano Zampini if (pcbddc->coarse_phi_D) { 170206656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 170306656605SStefano Zampini } 1704f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 170506656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 170606656605SStefano Zampini PetscScalar *marray; 170706656605SStefano Zampini 170806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 170906656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1710f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1711f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1712f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1713f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1714f4ddd8eeSStefano Zampini } 1715f4ddd8eeSStefano Zampini } 171606656605SStefano Zampini 1717f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 171806656605SStefano Zampini PetscScalar *marray; 171988ebb749SStefano Zampini 172006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17218eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 172206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 172388ebb749SStefano Zampini } 17243301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 172506656605SStefano Zampini n *= 2; 172688ebb749SStefano Zampini } 172706656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 172806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 172906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17308eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 173106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 173206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 173388ebb749SStefano Zampini } 17343301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 173506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17368eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 173706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 173806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 173988ebb749SStefano Zampini } 174088ebb749SStefano Zampini } else { 1741c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1742c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17431b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1744c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1745c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1746c0553b1fSStefano Zampini } 174788ebb749SStefano Zampini } 174806656605SStefano Zampini } 1749019a44ceSStefano Zampini 175006656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17514f1b2e48SStefano Zampini p0_lidx_I = NULL; 17524f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1753d12edf2fSStefano Zampini const PetscInt *idxs; 1754d12edf2fSStefano Zampini 1755d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17564f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17574f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17584f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17594f1b2e48SStefano Zampini } 1760d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1761d12edf2fSStefano Zampini } 1762d16cbb6bSStefano Zampini 176306656605SStefano Zampini /* vertices */ 176406656605SStefano Zampini if (n_vertices) { 176516f15bc4SStefano Zampini 1766*af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 176704708bb6SStefano Zampini 176816f15bc4SStefano Zampini if (n_R) { 176914393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 177006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 177116f15bc4SStefano Zampini PetscScalar *x,*y; 177204708bb6SStefano Zampini PetscBool isseqaij; 177306656605SStefano Zampini 177421eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 177514393ed6SStefano Zampini if (need_benign_correction) { 177614393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 177714393ed6SStefano Zampini IS is_p0; 177814393ed6SStefano Zampini PetscInt *idxs_p0,n; 177914393ed6SStefano Zampini 178014393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 178114393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 178214393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 1783*af25d912SStefano 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); 178414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 178514393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 178614393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 178714393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 178814393ed6SStefano Zampini } 178914393ed6SStefano Zampini 1790ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1791*af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1792ffd830a3SStefano Zampini } else { 1793ca92afb2SStefano Zampini PetscScalar *av,*array; 1794ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1795ca92afb2SStefano Zampini PetscInt n; 1796ca92afb2SStefano Zampini PetscBool flg_row; 1797ffd830a3SStefano Zampini 1798ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1799ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1800ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1801ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1802ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1803ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1804ca92afb2SStefano Zampini PetscInt j; 1805ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1806ffd830a3SStefano Zampini } 1807ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1808ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1809ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1810ffd830a3SStefano Zampini } 1811ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1812a3df083aSStefano Zampini if (need_benign_correction) { 1813df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1814a3df083aSStefano Zampini PetscScalar *marr; 1815a3df083aSStefano Zampini 1816a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 181714393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 181814393ed6SStefano Zampini 181914393ed6SStefano Zampini | 0 0 0 | (V) 182014393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 182114393ed6SStefano Zampini | 0 0 -1 | (p0) 182214393ed6SStefano Zampini 182314393ed6SStefano Zampini */ 1824df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 182514393ed6SStefano Zampini const PetscScalar *vals; 182614393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 182714393ed6SStefano Zampini PetscInt n,j,nz; 182814393ed6SStefano Zampini 1829df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1830df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 183114393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 183214393ed6SStefano Zampini for (j=0;j<n;j++) { 183314393ed6SStefano Zampini PetscScalar val = vals[j]; 183414393ed6SStefano Zampini PetscInt k,col = idxs[j]; 183514393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 183614393ed6SStefano Zampini } 183714393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1838df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 183914393ed6SStefano Zampini } 184072b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 184172b8c272SStefano Zampini } 184272b8c272SStefano Zampini if (F) { 184314393ed6SStefano Zampini /* need to correct the rhs */ 184472b8c272SStefano Zampini if (need_benign_correction) { 184572b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 184672b8c272SStefano Zampini PetscScalar *marr; 184772b8c272SStefano Zampini 184872b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 18495cbda25cSStefano Zampini if (lda_rhs != n_R) { 18505cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18515cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18525cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 18535cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18545cbda25cSStefano Zampini } 18555cbda25cSStefano Zampini } else { 1856a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1857a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18585cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1859a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1860a3df083aSStefano Zampini } 18615cbda25cSStefano Zampini } 1862a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1863a3df083aSStefano Zampini } 186406656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 186514393ed6SStefano Zampini /* need to correct the solution */ 1866a3df083aSStefano Zampini if (need_benign_correction) { 1867df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1868a3df083aSStefano Zampini PetscScalar *marr; 1869a3df083aSStefano Zampini 1870a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 18715cbda25cSStefano Zampini if (lda_rhs != n_R) { 18725cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18735cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18745cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 18755cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18765cbda25cSStefano Zampini } 18775cbda25cSStefano Zampini } else { 1878a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1879a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18805cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1881a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1882a3df083aSStefano Zampini } 18835cbda25cSStefano Zampini } 1884a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1885a3df083aSStefano Zampini } 188606656605SStefano Zampini } else { 188706656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 188806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1889ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1890ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 189106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 189206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 189306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 189406656605SStefano Zampini } 189506656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 189606656605SStefano Zampini } 189780677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1898ffd830a3SStefano Zampini /* S_VV and S_CV */ 189906656605SStefano Zampini if (n_constraints) { 190006656605SStefano Zampini Mat B; 190180677318SStefano Zampini 1902ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 190380677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1904ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1905ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 190680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 190780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 190880677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 190980677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 191080677318SStefano Zampini } 1911ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 191280677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 191380677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1914ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 191580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 191606656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1917ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1918ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 191906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 192006656605SStefano Zampini } 192104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 192204708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 1923511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 192404708bb6SStefano Zampini } 1925ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1926ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1927ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1928ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1929ffd830a3SStefano Zampini } 193006656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 193114393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 193214393ed6SStefano Zampini if (need_benign_correction) { 1933df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 193414393ed6SStefano Zampini PetscScalar *marr,*sums; 193514393ed6SStefano Zampini 193614393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 193714393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1938df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 193914393ed6SStefano Zampini const PetscScalar *vals; 194014393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 194114393ed6SStefano Zampini PetscInt n,j,nz; 194214393ed6SStefano Zampini 1943df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1944df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 194514393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 194614393ed6SStefano Zampini PetscInt k; 194714393ed6SStefano Zampini sums[j] = 0.; 194814393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 194914393ed6SStefano Zampini } 195014393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 195114393ed6SStefano Zampini for (j=0;j<n;j++) { 195214393ed6SStefano Zampini PetscScalar val = vals[j]; 195314393ed6SStefano Zampini PetscInt k; 195414393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 195514393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 195614393ed6SStefano Zampini } 195714393ed6SStefano Zampini } 195814393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1959df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 196014393ed6SStefano Zampini } 196114393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 196214393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 196314393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 196414393ed6SStefano Zampini } 196580677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 196606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 196706656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 196806656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 196906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 197006656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 197106656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 197206656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1973d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1974019a44ceSStefano Zampini } else { 1975d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1976d16cbb6bSStefano Zampini } 197721eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1978d16cbb6bSStefano Zampini 197906656605SStefano Zampini /* coarse basis functions */ 198006656605SStefano Zampini for (i=0;i<n_vertices;i++) { 198116f15bc4SStefano Zampini PetscScalar *y; 198216f15bc4SStefano Zampini 1983ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 198406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 198506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 198606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198806656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 198906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 199006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 199106656605SStefano Zampini 199206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19934f1b2e48SStefano Zampini PetscInt j; 19944f1b2e48SStefano Zampini 199506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 199606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 199706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20004f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 200106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 200206656605SStefano Zampini } 200306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 200406656605SStefano Zampini } 200504708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 200604708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 200706656605SStefano Zampini } 20085cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 200906656605SStefano Zampini 201006656605SStefano Zampini if (n_constraints) { 201106656605SStefano Zampini Mat B; 201206656605SStefano Zampini 2013ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 201406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 201580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 201606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 201706656605SStefano Zampini if (n_vertices) { 201880677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 201980677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 202080677318SStefano Zampini } else { 202180677318SStefano Zampini Mat S_VCt; 202280677318SStefano Zampini 2023ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2024ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 202572b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2026ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2027ffd830a3SStefano Zampini } 202880677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 202980677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 203080677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 203180677318SStefano Zampini } 203206656605SStefano Zampini } 203306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 203406656605SStefano Zampini /* coarse basis functions */ 203506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 203606656605SStefano Zampini PetscScalar *y; 203706656605SStefano Zampini 2038ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 203906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 204006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 204106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 204206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 204306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 204406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 204506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20464f1b2e48SStefano Zampini PetscInt j; 20474f1b2e48SStefano Zampini 204806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 204906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 205006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20534f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 205406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 205506656605SStefano Zampini } 205606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 205706656605SStefano Zampini } 205806656605SStefano Zampini } 205980677318SStefano Zampini if (n_constraints) { 206080677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 206180677318SStefano Zampini } 20624f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 206372b8c272SStefano Zampini 206472b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 206572b8c272SStefano Zampini if (pcbddc->benign_n) { 206672b8c272SStefano Zampini Mat B0_B,B0_BPHI; 206772b8c272SStefano Zampini IS is_dummy; 206872b8c272SStefano Zampini PetscScalar *data; 206972b8c272SStefano Zampini PetscInt j; 207072b8c272SStefano Zampini 207172b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 207272b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 207372b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 207472b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 207572b8c272SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 207672b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 207772b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 207872b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 207972b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 208072b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 208172b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 208272b8c272SStefano Zampini } 208372b8c272SStefano Zampini } 208472b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 208572b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 208672b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 208772b8c272SStefano Zampini } 2088019a44ceSStefano Zampini 208906656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 20903301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2091ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2092ffd830a3SStefano Zampini PetscScalar *marray; 209306656605SStefano Zampini 209406656605SStefano Zampini if (n_constraints) { 2095ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 209606656605SStefano Zampini 2097*af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 209806656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2099ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 210016f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 210106656605SStefano Zampini if (n_vertices) { 2102ffd830a3SStefano Zampini Mat S_VCT; 210306656605SStefano Zampini 210406656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2105ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 210616f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 210706656605SStefano Zampini } 2108ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 210906656605SStefano Zampini } 211016f15bc4SStefano Zampini if (n_vertices && n_R) { 2111ffd830a3SStefano Zampini PetscScalar *av,*marray; 2112ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2113ffd830a3SStefano Zampini PetscInt n; 2114ffd830a3SStefano Zampini PetscBool flg_row; 211506656605SStefano Zampini 2116ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2117*af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2118ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2119ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2120ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2121ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2122ffd830a3SStefano Zampini PetscInt j; 2123ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2124ffd830a3SStefano Zampini } 2125ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2126ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2127ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 212806656605SStefano Zampini } 212906656605SStefano Zampini 2130ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2131ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2132ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2133ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2134ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 213506656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 213606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 213706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 213806656605SStefano Zampini } 2139ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2140ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2141ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2142ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2143ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2144ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2145ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2146ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 214706656605SStefano Zampini } 2148ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 214906656605SStefano Zampini /* coarse basis functions */ 215006656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 215106656605SStefano Zampini PetscScalar *y; 215206656605SStefano Zampini 2153ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 215406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 215506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 215606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 215706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 215806656605SStefano Zampini if (i<n_vertices) { 215906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 216006656605SStefano Zampini } 216106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 216206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 216306656605SStefano Zampini 216406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 216506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 216606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 216706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 216806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 216906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 217006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 217106656605SStefano Zampini } 217206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217306656605SStefano Zampini } 2174ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2175ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 217606656605SStefano Zampini } 2177d62866d3SStefano Zampini /* free memory */ 217888ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 217906656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 218006656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 218106656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 218206656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2183d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2184d62866d3SStefano Zampini if (n_vertices) { 2185d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2186d62866d3SStefano Zampini } 2187d62866d3SStefano Zampini if (n_constraints) { 2188d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2189d62866d3SStefano Zampini } 219088ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 219188ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 219288ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2193d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 219488ebb749SStefano Zampini Mat coarse_sub_mat; 219525084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 219688ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 219788ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 219888ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 21998bec7fa6SStefano Zampini Mat C_B,CPHI; 22008bec7fa6SStefano Zampini IS is_dummy; 22018bec7fa6SStefano Zampini Vec mones; 220288ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 220388ebb749SStefano Zampini PetscReal real_value; 220488ebb749SStefano Zampini 2205a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2206a3df083aSStefano Zampini Mat A; 2207a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2208a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2209a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2210a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2211a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2212a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2213a3df083aSStefano Zampini } else { 221488ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 221588ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 221688ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 221788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2218a3df083aSStefano Zampini } 221988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 222088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2221ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 222288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 222388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 222488ebb749SStefano Zampini } 222588ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 222688ebb749SStefano Zampini 222725084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22283301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 222925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2230ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 223188ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 223288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 223388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 223488ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 223588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 223688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 223788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 223888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 223988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 224088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 224188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 224288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 224388ebb749SStefano Zampini } else { 224488ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 224588ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 224688ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 224788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 224888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 224988ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 225088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 225188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 225288ebb749SStefano Zampini } 225388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 225488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 225588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2256511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 22574f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2258fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2259d12edf2fSStefano Zampini PetscScalar *data,*data2; 22604f1b2e48SStefano Zampini PetscInt j; 2261d12edf2fSStefano Zampini 22624f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2263fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2264d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2265d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2266d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2267d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 22684f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 22694f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2270d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 22714f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 22724f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22734f1b2e48SStefano Zampini } 2274d12edf2fSStefano Zampini } 2275d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2276d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2277d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2278d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2279d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2280d12edf2fSStefano Zampini } 2281d12edf2fSStefano Zampini #if 0 2282d12edf2fSStefano Zampini { 2283d12edf2fSStefano Zampini PetscViewer viewer; 2284d12edf2fSStefano Zampini char filename[256]; 2285ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2286d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2287d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2288ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2289ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2290ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2291d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 229272b8c272SStefano Zampini if (save_change) { 229372b8c272SStefano Zampini Mat phi_B; 229472b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 229572b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 229672b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 229772b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 229872b8c272SStefano Zampini } else { 2299ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2300ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 230172b8c272SStefano Zampini } 2302ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2303ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2304ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2305ffd830a3SStefano Zampini } 2306ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2307ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2308ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2309ffd830a3SStefano Zampini } 231072b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2311ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2312ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2313ffd830a3SStefano Zampini } 2314d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2315d12edf2fSStefano Zampini } 2316d12edf2fSStefano Zampini #endif 231781d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23188bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23191575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 232006656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23218bec7fa6SStefano Zampini 23228bec7fa6SStefano Zampini /* check constraints */ 2323a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2324a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 23254f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23268bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2327a00504b5SStefano Zampini } else { 2328a00504b5SStefano Zampini PetscScalar *data; 2329a00504b5SStefano Zampini Mat tmat; 2330a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2331a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2332a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2333a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2334a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2335a00504b5SStefano Zampini } 23368bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23378bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23388bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23398bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2340bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2341ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2342bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2343bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2344bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2345bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2346bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 234788ebb749SStefano Zampini } 23488bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23498bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23508bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23518bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 235225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 235388ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 235488ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 235588ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 235688ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 235788ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 235888ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 235988ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 236088ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 236188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 236288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2363ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 236488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 236588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 236688ebb749SStefano Zampini } 236788ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 236888ebb749SStefano Zampini } 23698629588bSStefano Zampini /* get back data */ 23708629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 237188ebb749SStefano Zampini PetscFunctionReturn(0); 237288ebb749SStefano Zampini } 237388ebb749SStefano Zampini 237488ebb749SStefano Zampini #undef __FUNCT__ 2375d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2376d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2377aa0d41d4SStefano Zampini { 2378d65f70fdSStefano Zampini Mat *work_mat; 2379d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2380d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2381d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2382aa0d41d4SStefano Zampini PetscErrorCode ierr; 2383aa0d41d4SStefano Zampini 2384aa0d41d4SStefano Zampini PetscFunctionBegin; 2385d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2386d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2387d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2388d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2389aa0d41d4SStefano Zampini 2390d65f70fdSStefano Zampini if (!rsorted) { 2391906d46d4SStefano Zampini const PetscInt *idxs; 2392906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2393aa0d41d4SStefano Zampini 2394d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2395d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2396d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2397d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2398aa0d41d4SStefano Zampini } 2399d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2400d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2401d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2402d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2403aa0d41d4SStefano Zampini } 2404d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2405d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2406d65f70fdSStefano Zampini } else { 2407d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2408d65f70fdSStefano Zampini isrow_s = isrow; 2409aa0d41d4SStefano Zampini } 2410906d46d4SStefano Zampini 2411d65f70fdSStefano Zampini if (!csorted) { 2412d65f70fdSStefano Zampini if (isrow == iscol) { 2413d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2414d65f70fdSStefano Zampini iscol_s = isrow_s; 2415d65f70fdSStefano Zampini } else { 2416d65f70fdSStefano Zampini const PetscInt *idxs; 2417d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2418906d46d4SStefano Zampini 2419d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2420d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2421d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2422d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2423d65f70fdSStefano Zampini } 2424d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2425d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2426d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2427d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2428d65f70fdSStefano Zampini } 2429d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2430d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2431d65f70fdSStefano Zampini } 2432d65f70fdSStefano Zampini } else { 2433d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2434d65f70fdSStefano Zampini iscol_s = iscol; 2435d65f70fdSStefano Zampini } 2436d65f70fdSStefano Zampini 2437d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2438d65f70fdSStefano Zampini 2439d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2440906d46d4SStefano Zampini Mat new_mat; 2441d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2442906d46d4SStefano Zampini 2443d65f70fdSStefano Zampini if (!rsorted) { 2444d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2445d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2446d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2447d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2448906d46d4SStefano Zampini } 2449d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2450d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2451d65f70fdSStefano Zampini } else { 2452d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2453906d46d4SStefano Zampini } 2454d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2455d65f70fdSStefano Zampini 2456d65f70fdSStefano Zampini if (!csorted) { 2457d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2458d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2459d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2460d65f70fdSStefano Zampini } else { 2461d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2462d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2463d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2464d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2465d65f70fdSStefano Zampini } 2466d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2467d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2468d65f70fdSStefano Zampini } 2469d65f70fdSStefano Zampini } else { 2470d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2471d65f70fdSStefano Zampini } 2472d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2473d65f70fdSStefano Zampini 2474d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2475d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2476d65f70fdSStefano Zampini work_mat[0] = new_mat; 2477d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2478d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2479d65f70fdSStefano Zampini } 2480d65f70fdSStefano Zampini 2481d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2482d65f70fdSStefano Zampini *B = work_mat[0]; 2483d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2484d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2485d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2486d65f70fdSStefano Zampini PetscFunctionReturn(0); 2487d65f70fdSStefano Zampini } 2488d65f70fdSStefano Zampini 2489d65f70fdSStefano Zampini #undef __FUNCT__ 24905e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 24915e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2492aa0d41d4SStefano Zampini { 2493aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 24945e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2495d65f70fdSStefano Zampini Mat new_mat; 24965e8657edSStefano Zampini IS is_local,is_global; 2497d65f70fdSStefano Zampini PetscInt local_size; 2498d65f70fdSStefano Zampini PetscBool isseqaij; 2499aa0d41d4SStefano Zampini PetscErrorCode ierr; 2500aa0d41d4SStefano Zampini 2501aa0d41d4SStefano Zampini PetscFunctionBegin; 2502aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25035e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25045e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2505b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2506aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2507d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2508aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2509906d46d4SStefano Zampini 2510906d46d4SStefano Zampini /* check */ 2511906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2512906d46d4SStefano Zampini Vec x,x_change; 2513906d46d4SStefano Zampini PetscReal error; 2514906d46d4SStefano Zampini 25155e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2516906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25175e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2518e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2519e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2520d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2521e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2522e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2523906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2524906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2525906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2526906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2527906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2528906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2529906d46d4SStefano Zampini } 2530906d46d4SStefano Zampini 253122d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25329b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 253322d5777bSStefano Zampini if (isseqaij) { 2534a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2535a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2536aa0d41d4SStefano Zampini } else { 2537a00504b5SStefano Zampini Mat work_mat; 25381cf9b237SStefano Zampini 2539a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2540aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2541a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2542aa0d41d4SStefano Zampini } 25433301b35fSStefano Zampini if (matis->A->symmetric_set) { 25443301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2545e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25463301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2547e496cd5dSStefano Zampini #endif 25483301b35fSStefano Zampini } 2549d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2550aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2551aa0d41d4SStefano Zampini } 2552aa0d41d4SStefano Zampini 2553aa0d41d4SStefano Zampini #undef __FUNCT__ 2554a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25558ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2556a64d13efSStefano Zampini { 2557a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2558a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2559d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 256053892102SStefano Zampini PetscInt *idx_R_local=NULL; 25613a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25623a50541eSStefano Zampini PetscInt vbs,bs; 25636816873aSStefano Zampini PetscBT bitmask=NULL; 2564a64d13efSStefano Zampini PetscErrorCode ierr; 2565a64d13efSStefano Zampini 2566a64d13efSStefano Zampini PetscFunctionBegin; 2567b23d619eSStefano Zampini /* 2568b23d619eSStefano Zampini No need to setup local scatters if 2569b23d619eSStefano Zampini - primal space is unchanged 2570b23d619eSStefano Zampini AND 2571b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2572b23d619eSStefano Zampini AND 2573b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2574b23d619eSStefano Zampini */ 2575b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2576f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2577f4ddd8eeSStefano Zampini } 2578f4ddd8eeSStefano Zampini /* destroy old objects */ 2579f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2580f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2581f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2582a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2583b371cd4fSStefano Zampini n_B = pcis->n_B; 2584b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2585b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 25863a50541eSStefano Zampini 2587a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 25886816873aSStefano Zampini 258953892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2590df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2591854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2592a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2593a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25940e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2595a64d13efSStefano Zampini } 2596a64d13efSStefano Zampini 2597a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 25984641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 25996816873aSStefano Zampini idx_R_local[n_R++] = i; 2600a64d13efSStefano Zampini } 2601a64d13efSStefano Zampini } 2602df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2603df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26046816873aSStefano Zampini 2605df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2606df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26076816873aSStefano Zampini } 26083a50541eSStefano Zampini 26093a50541eSStefano Zampini /* Block code */ 26103a50541eSStefano Zampini vbs = 1; 26113a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26123a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26133a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26143a50541eSStefano Zampini PetscInt *vary; 2615df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2616785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26173a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2618d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2619d3df7717SStefano 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 */ 26200e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2621d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26223a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26233a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26243a50541eSStefano Zampini break; 26253a50541eSStefano Zampini } 26263a50541eSStefano Zampini } 2627d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2628d3df7717SStefano Zampini } else { 2629d3df7717SStefano Zampini /* Verify directly the R set */ 2630d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2631d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2632d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2633d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2634d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2635d3df7717SStefano Zampini break; 2636d3df7717SStefano Zampini } 2637d3df7717SStefano Zampini } 2638d3df7717SStefano Zampini } 2639d3df7717SStefano Zampini } 26403a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26413a50541eSStefano Zampini vbs = bs; 26423a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26433a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26443a50541eSStefano Zampini } 26453a50541eSStefano Zampini } 26463a50541eSStefano Zampini } 26473a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2648df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2649df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 265053892102SStefano Zampini 2651df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2652df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 265353892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2654df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 265553892102SStefano Zampini } else { 26563a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 265753892102SStefano Zampini } 2658a64d13efSStefano Zampini 2659a64d13efSStefano Zampini /* print some info if requested */ 2660a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2661a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2662a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26631575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2664a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2665a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 26664f1b2e48SStefano 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); 2667a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2668a64d13efSStefano Zampini } 2669a64d13efSStefano Zampini 2670a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2671df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 26726816873aSStefano Zampini IS is_aux1,is_aux2; 26736816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 26746816873aSStefano Zampini 26753a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2676854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2677854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2678a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26794641a718SStefano Zampini for (i=0; i<n_D; i++) { 26804641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26814641a718SStefano Zampini } 2682a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2683a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26844641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 26854641a718SStefano Zampini aux_array1[j++] = i; 2686a64d13efSStefano Zampini } 2687a64d13efSStefano Zampini } 2688a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2689a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2690a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 26914641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 26924641a718SStefano Zampini aux_array2[j++] = i; 2693a64d13efSStefano Zampini } 2694a64d13efSStefano Zampini } 2695a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2696a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2697a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2698a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2699a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2700a64d13efSStefano Zampini 27018eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2702785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2703a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27044641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27054641a718SStefano Zampini aux_array1[j++] = i; 2706a64d13efSStefano Zampini } 2707a64d13efSStefano Zampini } 2708a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2709a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2710a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2711a64d13efSStefano Zampini } 27124641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27133a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2714d62866d3SStefano Zampini } else { 2715df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27166816873aSStefano Zampini IS tis; 27176816873aSStefano Zampini PetscInt schur_size; 27186816873aSStefano Zampini 2719df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27206816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2721df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27226816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27236816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27246816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27256816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27266816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2727d62866d3SStefano Zampini } 2728d62866d3SStefano Zampini } 2729a64d13efSStefano Zampini PetscFunctionReturn(0); 2730a64d13efSStefano Zampini } 2731a64d13efSStefano Zampini 2732304d26faSStefano Zampini 2733304d26faSStefano Zampini #undef __FUNCT__ 2734304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2735684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2736304d26faSStefano Zampini { 2737304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2738304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2739304d26faSStefano Zampini PC pc_temp; 2740304d26faSStefano Zampini Mat A_RR; 2741f4ddd8eeSStefano Zampini MatReuse reuse; 2742304d26faSStefano Zampini PetscScalar m_one = -1.0; 2743304d26faSStefano Zampini PetscReal value; 274404708bb6SStefano Zampini PetscInt n_D,n_R; 27459577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2746304d26faSStefano Zampini PetscErrorCode ierr; 2747e604994aSStefano Zampini /* prefixes stuff */ 2748312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2749e604994aSStefano Zampini size_t len; 2750304d26faSStefano Zampini 2751304d26faSStefano Zampini PetscFunctionBegin; 2752304d26faSStefano Zampini 2753e604994aSStefano Zampini /* compute prefixes */ 2754e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2755e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2756e604994aSStefano Zampini if (!pcbddc->current_level) { 2757e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2758e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2759e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2760e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2761e604994aSStefano Zampini } else { 2762e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2763312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2764e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2765e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2766312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2767312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 276834d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 276934d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2770e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2771e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2772e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2773e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2774e604994aSStefano Zampini } 2775e604994aSStefano Zampini 2776304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2777684f6988SStefano Zampini if (dirichlet) { 2778d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2779450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2780450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2781450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2782450f8f5eSStefano Zampini } 2783450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2784a3df083aSStefano Zampini Mat A_IIn; 2785a3df083aSStefano Zampini 2786a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2787a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2788a3df083aSStefano Zampini pcis->A_II = A_IIn; 2789a3df083aSStefano Zampini } 2790450f8f5eSStefano Zampini } 27913301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27923301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2793964fefecSStefano Zampini } 2794ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2795964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2796304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2797304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2798304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2799304d26faSStefano Zampini /* default */ 2800304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2801e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28029577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2803304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28049577ea80SStefano Zampini if (issbaij) { 28059577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28069577ea80SStefano Zampini } else { 2807304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28089577ea80SStefano Zampini } 2809304d26faSStefano Zampini /* Allow user's customization */ 2810304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2811304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2812304d26faSStefano Zampini } 2813d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2814df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2815df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2816d62866d3SStefano Zampini 2817df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2818d5574798SStefano Zampini } 2819304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2820304d26faSStefano Zampini if (!n_D) { 2821304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2822304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2823304d26faSStefano Zampini } 2824304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2825304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2826304d26faSStefano Zampini /* set ksp_D into pcis data */ 2827304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2828304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2829304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2830684f6988SStefano Zampini } 2831304d26faSStefano Zampini 2832304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2833684f6988SStefano Zampini A_RR = 0; 2834684f6988SStefano Zampini if (neumann) { 2835d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 283604708bb6SStefano Zampini PetscInt ibs,mbs; 283704708bb6SStefano Zampini PetscBool issbaij; 283804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2839f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28408ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2841f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2842f4ddd8eeSStefano Zampini PetscInt nn_R; 284381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2844f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2845f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2846f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2847f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2848f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2849f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2850f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2851727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2852f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2853f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2854f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2855f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2856f4ddd8eeSStefano Zampini } 2857f4ddd8eeSStefano Zampini } 2858f4ddd8eeSStefano Zampini /* last check */ 2859d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2860f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2861f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2862f4ddd8eeSStefano Zampini } 2863f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2864f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2865f4ddd8eeSStefano Zampini } 2866a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 2867af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2868af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 286904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 287004708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 287104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 287204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 287304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2874af732b37SStefano Zampini } else { 2875511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 28766816873aSStefano Zampini } 287704708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 287804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 287904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 288004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 288104708bb6SStefano Zampini } else { 2882511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 288304708bb6SStefano Zampini } 288404708bb6SStefano Zampini } 2885a00504b5SStefano Zampini /* extract A_RR */ 2886a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 2887a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2888a00504b5SStefano Zampini 2889a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 289016e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2891a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 289216e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 289316e386b8SStefano Zampini } else { 2894a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 2895a00504b5SStefano Zampini } 2896a00504b5SStefano Zampini } else { 2897a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2898a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 2899a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2900a00504b5SStefano Zampini } 2901a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 2902f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 290316e386b8SStefano Zampini } 29043301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29053301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29066816873aSStefano Zampini } 2907f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2908304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2909304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2910304d26faSStefano Zampini /* default */ 2911304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2912e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2913304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29149577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29159577ea80SStefano Zampini if (issbaij) { 29169577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29179577ea80SStefano Zampini } else { 2918304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29199577ea80SStefano Zampini } 2920304d26faSStefano Zampini /* Allow user's customization */ 2921304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2922304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2923304d26faSStefano Zampini } 2924304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2925304d26faSStefano Zampini if (!n_R) { 2926304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2927304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2928304d26faSStefano Zampini } 29295cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2930df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2931df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2932df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2933d62866d3SStefano Zampini 2934df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2935d62866d3SStefano Zampini } 2936304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2937304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2938684f6988SStefano Zampini } 2939304d26faSStefano Zampini 2940304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29410fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2942684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2943684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29441575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2945684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2946684f6988SStefano Zampini } 2947684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29480fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29490fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29500fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29510fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29520fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2953304d26faSStefano Zampini /* need to be adapted? */ 2954b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2955b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2956b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2957304d26faSStefano Zampini /* print info */ 2958304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2959e604994aSStefano 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); 2960304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2961304d26faSStefano Zampini } 2962b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2963298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2964304d26faSStefano Zampini } 2965684f6988SStefano Zampini } 2966684f6988SStefano Zampini if (neumann) { /* Neumann */ 29670fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 29680fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 29690fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 29700fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 29710fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2972304d26faSStefano Zampini /* need to be adapted? */ 2973b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2974b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2975304d26faSStefano Zampini /* print info */ 2976304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2977e604994aSStefano 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); 2978304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2979304d26faSStefano Zampini } 2980b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2981298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2982304d26faSStefano Zampini } 29830fccc4e9SStefano Zampini } 2984684f6988SStefano Zampini } 29855cbda25cSStefano Zampini /* free Neumann problem's matrix */ 29865cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2987304d26faSStefano Zampini PetscFunctionReturn(0); 2988304d26faSStefano Zampini } 2989304d26faSStefano Zampini 2990304d26faSStefano Zampini #undef __FUNCT__ 2991ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 299280677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2993674ae819SStefano Zampini { 2994674ae819SStefano Zampini PetscErrorCode ierr; 2995674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2996be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2997674ae819SStefano Zampini 2998674ae819SStefano Zampini PetscFunctionBegin; 2999df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 300080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 300120c7b377SStefano Zampini } 300280677318SStefano Zampini if (!pcbddc->switch_static) { 300380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 300480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 300580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 300620c7b377SStefano Zampini } 3007df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 300880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 300980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301020c7b377SStefano Zampini } else { 3011df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3012be83ff47SStefano Zampini 3013df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3014df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 301520c7b377SStefano Zampini } 3016be83ff47SStefano Zampini } else { 301780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 302080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 302180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 302280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 302380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 302480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 302580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3026674ae819SStefano Zampini } 3027674ae819SStefano Zampini } 302872b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 302980677318SStefano Zampini if (applytranspose) { 303080677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 303180677318SStefano Zampini } else { 303280677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 303380677318SStefano Zampini } 3034be83ff47SStefano Zampini } else { 3035df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3036be83ff47SStefano Zampini 3037be83ff47SStefano Zampini if (applytranspose) { 3038df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3039be83ff47SStefano Zampini } else { 3040df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3041be83ff47SStefano Zampini } 3042be83ff47SStefano Zampini } 304380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 304480677318SStefano Zampini if (!pcbddc->switch_static) { 3045df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 304680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 304780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3048be83ff47SStefano Zampini } else { 3049df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3050be83ff47SStefano Zampini 3051df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3052df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3053be83ff47SStefano Zampini } 305480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 305580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 305680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 305780677318SStefano Zampini } 305880677318SStefano Zampini } else { 305980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 306480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 306580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 306680677318SStefano Zampini } 306780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 307080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3071674ae819SStefano Zampini } 3072674ae819SStefano Zampini PetscFunctionReturn(0); 3073674ae819SStefano Zampini } 3074674ae819SStefano Zampini 3075dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3076674ae819SStefano Zampini #undef __FUNCT__ 3077674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3078dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3079674ae819SStefano Zampini { 3080674ae819SStefano Zampini PetscErrorCode ierr; 3081674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3082674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3083674ae819SStefano Zampini const PetscScalar zero = 0.0; 3084674ae819SStefano Zampini 3085674ae819SStefano Zampini PetscFunctionBegin; 3086dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 30874fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3088dc359a40SStefano Zampini if (applytranspose) { 3089674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 30908eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3091dc359a40SStefano Zampini } else { 3092674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3093674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 309415aaf578SStefano Zampini } 30954fee134fSStefano Zampini } else { 30964fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 30974fee134fSStefano Zampini } 3098efc2fbd9SStefano Zampini 3099efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31004f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3101efc2fbd9SStefano Zampini PetscScalar *array; 31024f1b2e48SStefano Zampini PetscInt j; 3103efc2fbd9SStefano Zampini 3104efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31054f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3106efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3107efc2fbd9SStefano Zampini } 3108efc2fbd9SStefano Zampini 310912edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 311012edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 311112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 311212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 311312edc857SStefano Zampini 31149f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 311512edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 311612edc857SStefano Zampini if (pcbddc->coarse_ksp) { 311751694757SStefano Zampini Mat coarse_mat; 3118964fefecSStefano Zampini Vec rhs,sol; 311951694757SStefano Zampini MatNullSpace nullsp; 3120964fefecSStefano Zampini 3121964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3122964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 312351694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 312451694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 312551694757SStefano Zampini if (nullsp) { 312651694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 312751694757SStefano Zampini } 312812edc857SStefano Zampini if (applytranspose) { 3129964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 313012edc857SStefano Zampini } else { 3131964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 313212edc857SStefano Zampini } 313351694757SStefano Zampini if (nullsp) { 313451694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 313551694757SStefano Zampini } 313612edc857SStefano Zampini } 3137674ae819SStefano Zampini 3138674ae819SStefano Zampini /* Local solution on R nodes */ 31394fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 314080677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31419f00e9b4SStefano Zampini } 31429f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31439f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 314412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3145674ae819SStefano Zampini 31464fee134fSStefano Zampini /* Sum contributions from the two levels */ 31474fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3148dc359a40SStefano Zampini if (applytranspose) { 3149dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3150dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3151dc359a40SStefano Zampini } else { 3152674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31538eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3154dc359a40SStefano Zampini } 3155efc2fbd9SStefano Zampini /* store p0 */ 31564f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3157efc2fbd9SStefano Zampini PetscScalar *array; 31584f1b2e48SStefano Zampini PetscInt j; 3159efc2fbd9SStefano Zampini 3160efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31614f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3162efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3163efc2fbd9SStefano Zampini } 31644fee134fSStefano Zampini } else { /* expand the coarse solution */ 31654fee134fSStefano Zampini if (applytranspose) { 31664fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31674fee134fSStefano Zampini } else { 31684fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31694fee134fSStefano Zampini } 31704fee134fSStefano Zampini } 3171674ae819SStefano Zampini PetscFunctionReturn(0); 3172674ae819SStefano Zampini } 3173674ae819SStefano Zampini 3174674ae819SStefano Zampini #undef __FUNCT__ 3175674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 317612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3177674ae819SStefano Zampini { 3178674ae819SStefano Zampini PetscErrorCode ierr; 3179674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 318058da7f69SStefano Zampini PetscScalar *array; 318112edc857SStefano Zampini Vec from,to; 3182674ae819SStefano Zampini 3183674ae819SStefano Zampini PetscFunctionBegin; 318412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 318512edc857SStefano Zampini from = pcbddc->coarse_vec; 318612edc857SStefano Zampini to = pcbddc->vec1_P; 318712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 318812edc857SStefano Zampini Vec tvec; 318958da7f69SStefano Zampini 319058da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 319158da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 319212edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 319358da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 319458da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 319558da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 319612edc857SStefano Zampini } 319712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 319812edc857SStefano Zampini from = pcbddc->vec1_P; 319912edc857SStefano Zampini to = pcbddc->coarse_vec; 320012edc857SStefano Zampini } 320112edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3202674ae819SStefano Zampini PetscFunctionReturn(0); 3203674ae819SStefano Zampini } 3204674ae819SStefano Zampini 3205674ae819SStefano Zampini #undef __FUNCT__ 3206674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 320712edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3208674ae819SStefano Zampini { 3209674ae819SStefano Zampini PetscErrorCode ierr; 3210674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 321158da7f69SStefano Zampini PetscScalar *array; 321212edc857SStefano Zampini Vec from,to; 3213674ae819SStefano Zampini 3214674ae819SStefano Zampini PetscFunctionBegin; 321512edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 321612edc857SStefano Zampini from = pcbddc->coarse_vec; 321712edc857SStefano Zampini to = pcbddc->vec1_P; 321812edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 321912edc857SStefano Zampini from = pcbddc->vec1_P; 322012edc857SStefano Zampini to = pcbddc->coarse_vec; 322112edc857SStefano Zampini } 322212edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 322312edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 322412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 322512edc857SStefano Zampini Vec tvec; 322658da7f69SStefano Zampini 322712edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 322858da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 322958da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 323058da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 323158da7f69SStefano Zampini } 323258da7f69SStefano Zampini } else { 323358da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 323458da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 323512edc857SStefano Zampini } 323612edc857SStefano Zampini } 3237674ae819SStefano Zampini PetscFunctionReturn(0); 3238674ae819SStefano Zampini } 3239674ae819SStefano Zampini 3240984c4197SStefano Zampini /* uncomment for testing purposes */ 3241984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3242674ae819SStefano Zampini #undef __FUNCT__ 3243674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3244674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3245674ae819SStefano Zampini { 3246674ae819SStefano Zampini PetscErrorCode ierr; 3247674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3248674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3249674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3250984c4197SStefano Zampini /* one and zero */ 3251984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3252984c4197SStefano Zampini /* space to store constraints and their local indices */ 32539162d606SStefano Zampini PetscScalar *constraints_data; 32549162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32559162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32569162d606SStefano Zampini PetscInt *constraints_n; 3257984c4197SStefano Zampini /* iterators */ 3258b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3259984c4197SStefano Zampini /* BLAS integers */ 3260e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3261e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3262c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3263727cdba6SStefano Zampini /* reuse */ 32640e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 32650e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3266984c4197SStefano Zampini /* change of basis */ 3267b3d85658SStefano Zampini PetscBool qr_needed; 32689162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3269984c4197SStefano Zampini /* auxiliary stuff */ 327064efe560SStefano Zampini PetscInt *nnz,*is_indices; 32718a0068c3SStefano Zampini PetscInt ncc; 3272984c4197SStefano Zampini /* some quantities */ 327345a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3274a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3275984c4197SStefano Zampini 3276674ae819SStefano Zampini PetscFunctionBegin; 32778e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 32788e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 32798e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 328016909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3281088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3282088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 32830e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 32840e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 32850e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32860e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32870e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3288088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3289cf5a6209SStefano Zampini 3290cf5a6209SStefano Zampini /* print some info */ 3291cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3292cf5a6209SStefano Zampini IS vertices; 3293cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3294cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3295cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3296cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 32971575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3298cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3299cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3300fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3301fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3302cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33031575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3304cf5a6209SStefano Zampini } 3305cf5a6209SStefano Zampini 3306cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33079162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3308cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3309cf5a6209SStefano Zampini const Vec *nearnullvecs; 3310cf5a6209SStefano Zampini Vec *localnearnullsp; 3311cf5a6209SStefano Zampini PetscScalar *array; 3312cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3313cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3314674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3315b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3316674ae819SStefano Zampini PetscScalar *work; 3317674ae819SStefano Zampini PetscReal *singular_vals; 3318674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3319674ae819SStefano Zampini PetscReal *rwork; 3320674ae819SStefano Zampini #endif 3321674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3322674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3323674ae819SStefano Zampini #else 3324964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3325964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3326674ae819SStefano Zampini #endif 3327674ae819SStefano Zampini 3328674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3329d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3330d06fc5fdSStefano Zampini /* free unneeded index sets */ 3331d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3332d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3333674ae819SStefano Zampini } 3334d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3335d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3336d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3337d06fc5fdSStefano Zampini } 3338d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3339d06fc5fdSStefano Zampini n_ISForEdges = 0; 3340d06fc5fdSStefano Zampini } 3341d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3342d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3343d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3344d06fc5fdSStefano Zampini } 3345d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3346d06fc5fdSStefano Zampini n_ISForFaces = 0; 3347d06fc5fdSStefano Zampini } 334870022509SStefano Zampini 334970022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 335070022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 335170022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 335270022509SStefano Zampini if (pcbddc->NullSpace) { 335370022509SStefano Zampini PetscBool tbool[2],gbool[2]; 335470022509SStefano Zampini 335570022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3356b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3357d06fc5fdSStefano Zampini if (!ISForEdges) { 3358d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3359d06fc5fdSStefano Zampini } 3360b8ffe317SStefano Zampini } 3361d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3362d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3363b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3364d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3365d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 336698a51de6SStefano Zampini } 336770022509SStefano Zampini #endif 336808122e43SStefano Zampini 3369674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3370674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3371674ae819SStefano Zampini if (nearnullsp) { 3372674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3373f4ddd8eeSStefano Zampini /* remove any stored info */ 3374f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3375f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3376f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3377f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3378f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3379473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3380f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3381f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3382f4ddd8eeSStefano Zampini } 3383984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3384984c4197SStefano Zampini nnsp_size = 0; 3385674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3386674ae819SStefano Zampini } 3387984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3388984c4197SStefano Zampini max_constraints = nnsp_size; 3389984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3390984c4197SStefano Zampini 3391674ae819SStefano Zampini /* 3392674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 33939162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 33949162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 33959162d606SStefano Zampini There can be multiple constraints per connected component 3396674ae819SStefano Zampini */ 3397674ae819SStefano Zampini n_vertices = 0; 3398674ae819SStefano Zampini if (ISForVertices) { 3399674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3400674ae819SStefano Zampini } 34019162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34029162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34039162d606SStefano Zampini 34049162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34059162d606SStefano Zampini total_counts *= max_constraints; 3406674ae819SStefano Zampini total_counts += n_vertices; 34074641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34089162d606SStefano Zampini 3409674ae819SStefano Zampini total_counts = 0; 3410674ae819SStefano Zampini max_size_of_constraint = 0; 3411674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34129162d606SStefano Zampini IS used_is; 3413674ae819SStefano Zampini if (i<n_ISForEdges) { 34149162d606SStefano Zampini used_is = ISForEdges[i]; 3415674ae819SStefano Zampini } else { 34169162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3417674ae819SStefano Zampini } 34189162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3419674ae819SStefano Zampini total_counts += j; 3420674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3421674ae819SStefano Zampini } 34229162d606SStefano 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); 34239162d606SStefano Zampini 3424984c4197SStefano Zampini /* get local part of global near null space vectors */ 3425785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3426984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3427984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3428e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3429e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3430984c4197SStefano Zampini } 3431674ae819SStefano Zampini 3432242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3433242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3434a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3435242a89d7SStefano Zampini 3436984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3437a773dcb8SStefano Zampini if (!skip_lapack) { 3438674ae819SStefano Zampini PetscScalar temp_work; 3439911cabfeSStefano Zampini 3440674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3441984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3442785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3443785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3444785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3445674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3446785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3447674ae819SStefano Zampini #endif 3448674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3449c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3450c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3451674ae819SStefano Zampini lwork = -1; 3452674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3453674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3454c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3455674ae819SStefano Zampini #else 3456c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3457674ae819SStefano Zampini #endif 3458674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3459984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3460674ae819SStefano Zampini #else /* on missing GESVD */ 3461674ae819SStefano Zampini /* SVD */ 3462674ae819SStefano Zampini PetscInt max_n,min_n; 3463674ae819SStefano Zampini max_n = max_size_of_constraint; 3464984c4197SStefano Zampini min_n = max_constraints; 3465984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3466674ae819SStefano Zampini min_n = max_size_of_constraint; 3467984c4197SStefano Zampini max_n = max_constraints; 3468674ae819SStefano Zampini } 3469785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3470674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3471785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3472674ae819SStefano Zampini #endif 3473674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3474674ae819SStefano Zampini lwork = -1; 3475e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3476e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3477b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3478674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3479674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 34809162d606SStefano 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)); 3481674ae819SStefano Zampini #else 34829162d606SStefano 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)); 3483674ae819SStefano Zampini #endif 3484674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3485984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3486984c4197SStefano Zampini #endif /* on missing GESVD */ 3487674ae819SStefano Zampini /* Allocate optimal workspace */ 3488674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3489854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3490674ae819SStefano Zampini } 3491674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3492674ae819SStefano Zampini total_counts = 0; 34939162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 34949162d606SStefano Zampini constraints_data_ptr[0] = 0; 3495674ae819SStefano Zampini /* vertices */ 34969162d606SStefano Zampini if (n_vertices) { 3497674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 34989162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3499674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35009162d606SStefano Zampini constraints_n[total_counts] = 1; 35019162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35029162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35039162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3504674ae819SStefano Zampini total_counts++; 3505674ae819SStefano Zampini } 3506674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3507674ae819SStefano Zampini n_vertices = total_counts; 3508674ae819SStefano Zampini } 3509984c4197SStefano Zampini 3510674ae819SStefano Zampini /* edges and faces */ 35119162d606SStefano Zampini total_counts_cc = total_counts; 3512911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35139162d606SStefano Zampini IS used_is; 35149162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35159162d606SStefano Zampini 3516911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35179162d606SStefano Zampini used_is = ISForEdges[ncc]; 3518984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3519674ae819SStefano Zampini } else { 35209162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3521984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3522674ae819SStefano Zampini } 3523674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35249162d606SStefano Zampini 35259162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35269162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3527984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3528984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3529674ae819SStefano Zampini if (nnsp_has_cnst) { 35305b08dc53SStefano Zampini PetscScalar quad_value; 35319162d606SStefano Zampini 35329162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35339162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35349162d606SStefano Zampini 3535a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3536674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3537a773dcb8SStefano Zampini } else { 3538a773dcb8SStefano Zampini quad_value = 1.0; 3539a773dcb8SStefano Zampini } 3540674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35419162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3542674ae819SStefano Zampini } 35439162d606SStefano Zampini temp_constraints++; 3544674ae819SStefano Zampini total_counts++; 3545674ae819SStefano Zampini } 3546674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3547984c4197SStefano Zampini PetscReal real_value; 35489162d606SStefano Zampini PetscScalar *ptr_to_data; 35499162d606SStefano Zampini 3550984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35519162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3552674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35539162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3554674ae819SStefano Zampini } 3555984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3556984c4197SStefano Zampini /* check if array is null on the connected component */ 3557e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35589162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35595b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3560674ae819SStefano Zampini temp_constraints++; 3561674ae819SStefano Zampini total_counts++; 35629162d606SStefano Zampini if (!idxs_copied) { 35639162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35649162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3565674ae819SStefano Zampini } 3566674ae819SStefano Zampini } 35679162d606SStefano Zampini } 35689162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 356945a1bb75SStefano Zampini valid_constraints = temp_constraints; 3570eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3571a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 35729162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 35739162d606SStefano Zampini 35749162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3575a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35769162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3577a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 35789162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3579a773dcb8SStefano Zampini } else { /* perform SVD */ 3580984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 35819162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3582674ae819SStefano Zampini 3583674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3584984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3585984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3586984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3587984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3588984c4197SStefano Zampini from that computed using LAPACKgesvd 3589984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3590984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3591984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3592674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3593e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3594984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3595674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3596674ae819SStefano Zampini for (k=0;k<j+1;k++) { 35979162d606SStefano 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)); 3598674ae819SStefano Zampini } 3599674ae819SStefano Zampini } 3600e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3601e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3602e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3603674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3604c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3605674ae819SStefano Zampini #else 3606c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3607674ae819SStefano Zampini #endif 3608674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3609984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3610984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3611674ae819SStefano Zampini j = 0; 3612984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3613674ae819SStefano Zampini total_counts = total_counts-j; 361445a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3615e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3616c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3617c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3618c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3619c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3620c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3621c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3622674ae819SStefano Zampini if (j<temp_constraints) { 3623984c4197SStefano Zampini PetscInt ii; 3624984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3625674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36269162d606SStefano 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)); 3627674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3628984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3629674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36309162d606SStefano 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]; 3631674ae819SStefano Zampini } 3632674ae819SStefano Zampini } 3633674ae819SStefano Zampini } 3634674ae819SStefano Zampini #else /* on missing GESVD */ 3635e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3636e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3637b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3638674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3639674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36409162d606SStefano 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)); 3641674ae819SStefano Zampini #else 36429162d606SStefano 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)); 3643674ae819SStefano Zampini #endif 3644984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3645674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3646984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3647e310c8b4SStefano Zampini k = temp_constraints; 3648e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3649674ae819SStefano Zampini j = 0; 3650e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 365145a1bb75SStefano Zampini valid_constraints = k-j; 3652911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3653984c4197SStefano Zampini #endif /* on missing GESVD */ 3654674ae819SStefano Zampini } 3655a773dcb8SStefano Zampini } 36569162d606SStefano Zampini /* update pointers information */ 36579162d606SStefano Zampini if (valid_constraints) { 36589162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36599162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36609162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36619162d606SStefano Zampini /* set change_of_basis flag */ 366245a1bb75SStefano Zampini if (boolforchange) { 3663b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 36649162d606SStefano Zampini } 3665b3d85658SStefano Zampini total_counts_cc++; 366645a1bb75SStefano Zampini } 366745a1bb75SStefano Zampini } 3668984c4197SStefano Zampini /* free workspace */ 36698f1c130eSStefano Zampini if (!skip_lapack) { 3670984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3671984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3672984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3673984c4197SStefano Zampini #endif 3674984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3675984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3676984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3677984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3678984c4197SStefano Zampini #endif 3679984c4197SStefano Zampini } 3680984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3681984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3682984c4197SStefano Zampini } 3683984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3684cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3685cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3686cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3687cf5a6209SStefano Zampini } 3688cf5a6209SStefano Zampini if (n_ISForFaces) { 3689cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3690cf5a6209SStefano Zampini } 3691cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3692cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3693cf5a6209SStefano Zampini } 3694cf5a6209SStefano Zampini if (n_ISForEdges) { 3695cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3696cf5a6209SStefano Zampini } 3697cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 369808122e43SStefano Zampini } else { 369908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3700984c4197SStefano Zampini 370108122e43SStefano Zampini total_counts = 0; 370208122e43SStefano Zampini n_vertices = 0; 3703d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3704d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 370508122e43SStefano Zampini } 370608122e43SStefano Zampini max_constraints = 0; 37079162d606SStefano Zampini total_counts_cc = 0; 370808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 370908122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37109162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 371108122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 371208122e43SStefano Zampini } 37139162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37149162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37159162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37169162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 371774d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37189162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37199162d606SStefano Zampini total_counts_cc = 0; 37209162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37219162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37229162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 372308122e43SStefano Zampini } 372408122e43SStefano Zampini } 37259162d606SStefano Zampini #if 0 37269162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37279162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37289162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37299162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37309162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37319162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37329162d606SStefano Zampini } 37339162d606SStefano Zampini printf("\n"); 37349162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37359162d606SStefano Zampini } 37361b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37378bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37381b968477SStefano Zampini } 37391b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37408bec7fa6SStefano 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]); 37411b968477SStefano Zampini } 374208122e43SStefano Zampini #endif 374308122e43SStefano Zampini 37448bec7fa6SStefano Zampini max_size_of_constraint = 0; 37459162d606SStefano 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]); 37469162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 374708122e43SStefano Zampini /* Change of basis */ 3748b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 374908122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 375008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 375108122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3752b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 375308122e43SStefano Zampini } 375408122e43SStefano Zampini } 375508122e43SStefano Zampini } 375608122e43SStefano Zampini } 3757984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37584f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 375908122e43SStefano Zampini 37609162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37619162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 37626c4ed002SBarry 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); 3763674ae819SStefano Zampini 3764674ae819SStefano Zampini /* Create constraint matrix */ 3765674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 376616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3767984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3768984c4197SStefano Zampini 3769984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3770a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3771a717540cSStefano Zampini qr_needed = PETSC_FALSE; 377274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3773984c4197SStefano Zampini total_primal_vertices=0; 3774b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 37759162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37769162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 377772b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 37789162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3779b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 378064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 37819162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37829162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3783a717540cSStefano Zampini } 3784b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 378574d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3786a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3787a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3788a717540cSStefano Zampini } 3789fa434743SStefano Zampini } else { 3790b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3791fa434743SStefano Zampini } 3792a717540cSStefano Zampini } 3793b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3794b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3795674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 379670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3797b3d85658SStefano Zampini 37984f1b2e48SStefano 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); 37990e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38000e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3801984c4197SStefano Zampini 3802984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 380374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3804785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3805984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 380674d5cdf7SStefano Zampini 3807984c4197SStefano Zampini j = total_primal_vertices; 380874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3809b3d85658SStefano Zampini cum = total_primal_vertices; 38109162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38114641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3812b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3813b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3814b3d85658SStefano Zampini cum++; 38159162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 381674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 381774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 381874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 381974d5cdf7SStefano Zampini } 38209162d606SStefano Zampini j += constraints_n[i]; 3821674ae819SStefano Zampini } 3822674ae819SStefano Zampini } 3823674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3824674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3825088faed8SStefano Zampini 3826674ae819SStefano Zampini /* set values in constraint matrix */ 3827984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38280e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3829674ae819SStefano Zampini } 3830984c4197SStefano Zampini total_counts = total_primal_vertices; 38319162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38324641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38339162d606SStefano Zampini PetscInt *cols; 38349162d606SStefano Zampini 38359162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38369162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38379162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38389162d606SStefano Zampini PetscInt row = total_counts+k; 38399162d606SStefano Zampini PetscScalar *vals; 38409162d606SStefano Zampini 38419162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38429162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38439162d606SStefano Zampini } 38449162d606SStefano Zampini total_counts += constraints_n[i]; 3845674ae819SStefano Zampini } 3846674ae819SStefano Zampini } 3847674ae819SStefano Zampini /* assembling */ 3848674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3849674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3850088faed8SStefano Zampini 3851984c4197SStefano Zampini /* 38526a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3853984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3854f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 3855984c4197SStefano Zampini */ 3856674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3857674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3858026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3859984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3860984c4197SStefano Zampini /* working stuff for GEQRF */ 386181d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3862984c4197SStefano Zampini PetscBLASInt lqr_work; 3863984c4197SStefano Zampini /* working stuff for UNGQR */ 3864984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3865984c4197SStefano Zampini PetscBLASInt lgqr_work; 3866984c4197SStefano Zampini /* working stuff for TRTRS */ 3867984c4197SStefano Zampini PetscScalar *trs_rhs; 38683f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3869984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3870984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3871984c4197SStefano Zampini PetscScalar *start_vals; 3872984c4197SStefano Zampini /* working stuff for values insertion */ 38734641a718SStefano Zampini PetscBT is_primal; 387464efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3875906d46d4SStefano Zampini /* matrix sizes */ 3876906d46d4SStefano Zampini PetscInt global_size,local_size; 3877906d46d4SStefano Zampini /* temporary change of basis */ 3878906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3879cf5a6209SStefano Zampini /* extra space for debugging */ 3880cf5a6209SStefano Zampini PetscScalar *dbg_work; 3881984c4197SStefano Zampini 3882906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3883906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 388416f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3885bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3886906d46d4SStefano Zampini /* nonzeros for local mat */ 3887bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 38881dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3889bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 38901dd7afcfSStefano Zampini } else { 38911dd7afcfSStefano Zampini const PetscInt *ii; 38921dd7afcfSStefano Zampini PetscInt n; 38931dd7afcfSStefano Zampini PetscBool flg_row; 38941dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 38951dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 38961dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 38971dd7afcfSStefano Zampini } 38989162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3899a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39009162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3901a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39029162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3903a717540cSStefano Zampini } else { 39049162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39059162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3906a717540cSStefano Zampini } 3907a717540cSStefano Zampini } 3908a717540cSStefano Zampini } 3909906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3910bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 39111dd7afcfSStefano Zampini /* Set interior change in the matrix */ 39121dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3913bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3914906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3915a717540cSStefano Zampini } 39161dd7afcfSStefano Zampini } else { 39171dd7afcfSStefano Zampini const PetscInt *ii,*jj; 39181dd7afcfSStefano Zampini PetscScalar *aa; 39191dd7afcfSStefano Zampini PetscInt n; 39201dd7afcfSStefano Zampini PetscBool flg_row; 39211dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39221dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39231dd7afcfSStefano Zampini for (i=0;i<n;i++) { 39241dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 39251dd7afcfSStefano Zampini } 39261dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39271dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39281dd7afcfSStefano Zampini } 3929a717540cSStefano Zampini 3930a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3931a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3932a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3933a717540cSStefano Zampini } 3934a717540cSStefano Zampini 3935a717540cSStefano Zampini 3936a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3937a717540cSStefano Zampini /* 3938a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3939a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3940a717540cSStefano Zampini 3941a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3942a717540cSStefano Zampini 3943a6b551f4SStefano 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) 3944a6b551f4SStefano Zampini 3945a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3946a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3947a717540cSStefano Zampini | ... | 3948a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3949a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3950a717540cSStefano Zampini 3951a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3952a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3953a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3954a6b551f4SStefano Zampini 3955a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3956a717540cSStefano Zampini */ 3957a717540cSStefano Zampini if (qr_needed) { 3958984c4197SStefano Zampini /* space to store Q */ 3959854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3960984c4197SStefano Zampini /* first we issue queries for optimal work */ 39613f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39623f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 39633f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3964984c4197SStefano Zampini lqr_work = -1; 39653f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3966984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3967984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3968785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3969984c4197SStefano Zampini lgqr_work = -1; 39703f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39713f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 39723f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 39733f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39743f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 39753f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3976984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3977984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3978785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3979984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3980785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3981984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3982785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3983a717540cSStefano Zampini /* allocating workspace for check */ 3984a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3985cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3986a717540cSStefano Zampini } 3987a717540cSStefano Zampini } 3988984c4197SStefano Zampini /* array to store whether a node is primal or not */ 39894641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3990473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 39910e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 39926c4ed002SBarry 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); 399339e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 399439e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 399539e2fb2aSStefano Zampini } 399639e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3997984c4197SStefano Zampini 3998a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 39999162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40009162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40014641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4002984c4197SStefano Zampini /* get constraint info */ 40039162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4004984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4005984c4197SStefano Zampini 4006984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40079162d606SStefano 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); 4008674ae819SStefano Zampini } 4009984c4197SStefano Zampini 4010fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4011a717540cSStefano Zampini 4012a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4013a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40149162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4015a717540cSStefano Zampini } 4016984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40179162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4018984c4197SStefano Zampini 4019984c4197SStefano Zampini /* compute QR decomposition of constraints */ 40203f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40213f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4023674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40243f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4025984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4026674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4027984c4197SStefano Zampini 4028984c4197SStefano Zampini /* explictly compute R^-T */ 4029984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4030984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40313f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40323f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40333f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40343f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4035984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40363f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4037984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4038984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4039984c4197SStefano Zampini 4040a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40413f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40423f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40433f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40443f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4045984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40463f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4047984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4048984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4049984c4197SStefano Zampini 4050984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4051984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4052984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40533f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40543f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40553f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40563f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40573f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4059984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40609162d606SStefano 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)); 4061984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40629162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4063984c4197SStefano Zampini 4064984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 40659162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4066984c4197SStefano Zampini /* insert cols for primal dofs */ 4067984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4068984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 40699162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4070906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4071984c4197SStefano Zampini } 4072984c4197SStefano Zampini /* insert cols for dual dofs */ 4073984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 40749162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4075984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 40769162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4077906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4078984c4197SStefano Zampini j++; 4079674ae819SStefano Zampini } 4080674ae819SStefano Zampini } 4081984c4197SStefano Zampini 4082984c4197SStefano Zampini /* check change of basis */ 4083984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4084984c4197SStefano Zampini PetscInt ii,jj; 4085984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4086c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4087c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4088c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4089c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4090c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4091c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4092984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4093cf5a6209SStefano 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)); 4094984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4095984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4096984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4097cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4098cf5a6209SStefano 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; 4099674ae819SStefano Zampini } 4100674ae819SStefano Zampini } 4101984c4197SStefano Zampini if (!valid_qr) { 410222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4103984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4104984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4105cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4106cf5a6209SStefano 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])); 4107674ae819SStefano Zampini } 4108cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4109cf5a6209SStefano 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])); 4110984c4197SStefano Zampini } 4111984c4197SStefano Zampini } 4112984c4197SStefano Zampini } 4113674ae819SStefano Zampini } else { 411422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4115674ae819SStefano Zampini } 4116674ae819SStefano Zampini } 4117a717540cSStefano Zampini } else { /* simple transformation block */ 4118a717540cSStefano Zampini PetscInt row,col; 4119a6b551f4SStefano Zampini PetscScalar val,norm; 4120a6b551f4SStefano Zampini 4121a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41229162d606SStefano 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)); 4123a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 41249162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 41259162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4126bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 41279162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4128906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41299162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4130a717540cSStefano Zampini } else { 4131a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41329162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4133a717540cSStefano Zampini if (row != col) { 41349162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4135a717540cSStefano Zampini } else { 41369162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4137a717540cSStefano Zampini } 4138906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4139a717540cSStefano Zampini } 4140a717540cSStefano Zampini } 4141a717540cSStefano Zampini } 414298a51de6SStefano Zampini if (pcbddc->dbg_flag) { 414322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4144a717540cSStefano Zampini } 4145674ae819SStefano Zampini } 4146984c4197SStefano Zampini } else { 4147984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41489162d606SStefano 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); 4149674ae819SStefano Zampini } 4150674ae819SStefano Zampini } 4151674ae819SStefano Zampini } 4152a717540cSStefano Zampini 4153a717540cSStefano Zampini /* free workspace */ 4154a717540cSStefano Zampini if (qr_needed) { 4155984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4156cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4157984c4197SStefano Zampini } 4158984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4159984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4160984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4161984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4162984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4163674ae819SStefano Zampini } 4164a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4165906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4166906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4167906d46d4SStefano Zampini 4168906d46d4SStefano Zampini /* assembling of global change of variable */ 416988c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4170bbb9e6c6SStefano Zampini Mat tmat; 417116f15bc4SStefano Zampini PetscInt bs; 417216f15bc4SStefano Zampini 4173906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4174906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4175bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4176bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4177bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4178bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 417916f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 418016f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4181906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4182bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4183bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4184bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4185bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4186bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4187e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4188e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4189bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4190bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 419188c03ad3SStefano Zampini 4192906d46d4SStefano Zampini /* check */ 4193906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4194906d46d4SStefano Zampini PetscReal error; 4195906d46d4SStefano Zampini Vec x,x_change; 4196906d46d4SStefano Zampini 4197906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4198906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4199906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4200906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4201e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4202e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4203bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4204e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4205e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4206906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4207906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4208906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4209906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4210bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4211906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4212906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4213906d46d4SStefano Zampini } 4214b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4215b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4216b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4217bf3a8328SStefano Zampini 4218bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4219bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4220bf3a8328SStefano Zampini } 4221b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4222ac632422SStefano Zampini Mat S_new,tmat; 4223bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4224bbb9e6c6SStefano Zampini 4225bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 42266816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4227bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4228bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4229bf3a8328SStefano Zampini IS is_V; 4230b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4231b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4232b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4233b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4234b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4235bf3a8328SStefano Zampini } 4236bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4237ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4238b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4239ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4240bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4241bf3a8328SStefano Zampini const PetscScalar *array; 4242bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4243bf3a8328SStefano Zampini PetscInt i,n_V; 4244bf3a8328SStefano Zampini 4245b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4246b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4247b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4248b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4249b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4250b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4251b087196eSStefano Zampini PetscScalar val; 4252b087196eSStefano Zampini PetscInt idx; 4253b087196eSStefano Zampini 4254b087196eSStefano Zampini idx = idxs_V[i]; 4255b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4256b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4257b087196eSStefano Zampini } 4258b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4259b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4260bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4261bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4262bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4263bf3a8328SStefano Zampini } 4264ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4265ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4266ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4267ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4268b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4269ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4270bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4271b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4272bf3a8328SStefano Zampini } 4273ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4274ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4275ac632422SStefano Zampini } 4276b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 427788c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4278b96c3477SStefano Zampini } 4279c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4280c9db6a07SStefano Zampini if (sub_schurs->change) { 4281c9db6a07SStefano Zampini PetscInt i; 4282c9db6a07SStefano Zampini 4283c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4284c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4285c9db6a07SStefano Zampini } 4286c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4287c9db6a07SStefano Zampini } 4288b96c3477SStefano Zampini } 428916909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 429016909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 429116909a7fSStefano Zampini } else { 4292906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 429316909a7fSStefano Zampini } 42941dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 42951dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 42961dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 42971dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->change_interior,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 42981dd7afcfSStefano Zampini } 429972b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 430072b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 430172b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 430272b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 430372b8c272SStefano Zampini } 43041dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 43051dd7afcfSStefano Zampini PetscBool needglobal = PETSC_FALSE; 43061dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43071dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43081dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&needglobal,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43091dd7afcfSStefano Zampini } 43101dd7afcfSStefano Zampini if (!needglobal && pcbddc->user_ChangeOfBasisMatrix) { 4311b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4312b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4313906d46d4SStefano Zampini } else { 43141dd7afcfSStefano Zampini Mat benign_global = NULL; 43151dd7afcfSStefano Zampini if (needglobal) { 43161dd7afcfSStefano Zampini Mat tmat; 43171dd7afcfSStefano Zampini 43181dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 43191dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 43201dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 43211dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43221dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43231dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 43241dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43251dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43261dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 43271dd7afcfSStefano Zampini if (pcbddc->benign_change) { 43281dd7afcfSStefano Zampini Mat M; 43291dd7afcfSStefano Zampini 43301dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 43311dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 43321dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 43331dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4334906d46d4SStefano Zampini } else { 43351dd7afcfSStefano Zampini Mat eye; 43361dd7afcfSStefano Zampini PetscScalar *array; 43371dd7afcfSStefano Zampini 43381dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43391dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 43401dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 43411dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4342906d46d4SStefano Zampini } 43431dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43441dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43451dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43461dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 43471dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 43481dd7afcfSStefano Zampini } 43491dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 43501dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 43511dd7afcfSStefano Zampini } 43521dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 43531dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 43541dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 43551dd7afcfSStefano Zampini } else if (needglobal) { 43561dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 43571dd7afcfSStefano Zampini } 43581dd7afcfSStefano Zampini } 435916909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 436016909a7fSStefano Zampini IS is_global; 436116909a7fSStefano Zampini const PetscInt *gidxs; 436216909a7fSStefano Zampini 436316909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 436416909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 436516909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 436616909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 436716909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 436816909a7fSStefano Zampini } 43691dd7afcfSStefano Zampini } 43701dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 43711dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4372b9b85e73SStefano Zampini } 4373a717540cSStefano Zampini 437472b8c272SStefano Zampini if (!pcbddc->fake_change) { 43754f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 43764f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 43774f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 43784f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4379019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4380019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4381019a44ceSStefano Zampini pcbddc->local_primal_size++; 4382019a44ceSStefano Zampini } 4383019a44ceSStefano Zampini 4384019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4385727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4386727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 43879f47a83aSStefano 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); 4388c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 43890e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 43909f47a83aSStefano 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); 4391727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4392727cdba6SStefano Zampini } 43930e6343abSStefano Zampini } 4394727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4395b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 439672b8c272SStefano Zampini } 439772b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4398727cdba6SStefano Zampini 4399a717540cSStefano Zampini /* flush dbg viewer */ 4400b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4401b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4402b8ffe317SStefano Zampini } 4403a717540cSStefano Zampini 4404e310c8b4SStefano Zampini /* free workspace */ 4405a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 44064641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 440708122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 44089162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 44099162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 441008122e43SStefano Zampini } else { 44119162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 44129162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 44139162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 441408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 441508122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 44169162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 44179162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 441808122e43SStefano Zampini } 4419674ae819SStefano Zampini PetscFunctionReturn(0); 4420674ae819SStefano Zampini } 4421674ae819SStefano Zampini 4422674ae819SStefano Zampini #undef __FUNCT__ 4423674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4424674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4425674ae819SStefano Zampini { 4426674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4427674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4428674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 442914f95afaSStefano Zampini PetscInt ierr,i,N; 4430674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4431674ae819SStefano Zampini 4432674ae819SStefano Zampini PetscFunctionBegin; 44338e61c736SStefano Zampini /* Reset previously computed graph */ 44348e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4435674ae819SStefano Zampini /* Init local Graph struct */ 44367fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 44373bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4438674ae819SStefano Zampini 4439575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 44405099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 44415099eff2SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid size of local CSR graph! Found %d, expected %d\n",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs); 4442575ad6abSStefano Zampini } 44439577ea80SStefano Zampini 4444674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4445d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 44464d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 44474d379d7bSStefano Zampini PetscInt nvtxs; 4448e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4449674ae819SStefano Zampini 44502fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 44512fffb893SStefano Zampini if (flg_row) { 44524d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4453b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 44542fffb893SStefano Zampini } 44552fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4456674ae819SStefano Zampini } 44579b28b941SStefano Zampini if (pcbddc->dbg_flag) { 44589b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4459674ae819SStefano Zampini } 4460674ae819SStefano Zampini 446163602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 446263602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 446363602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 446495ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 446563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4466e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 446763602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4468674ae819SStefano Zampini } 446963602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 447063602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 447163602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4472674ae819SStefano Zampini } 447363602bcaSStefano Zampini } else { 447463602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 447563602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4476854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 447763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 447863602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 447963602bcaSStefano Zampini } 448063602bcaSStefano Zampini } 4481674ae819SStefano Zampini } 4482674ae819SStefano Zampini 4483674ae819SStefano Zampini /* Setup of Graph */ 4484785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4485e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4486785d1243SStefano Zampini } 4487785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4488e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4489785d1243SStefano Zampini } 449030368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 449130368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 449230368db7SStefano Zampini } 449314f95afaSStefano 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); 4494674ae819SStefano Zampini 44954f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 44964f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 44974f1b2e48SStefano Zampini PetscInt *local_subs; 44984f1b2e48SStefano Zampini 44994f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 45004f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 45014f1b2e48SStefano Zampini const PetscInt *idxs; 45024f1b2e48SStefano Zampini PetscInt nl,j; 45034f1b2e48SStefano Zampini 45044f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 45054f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45064f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 45074f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 45084f1b2e48SStefano Zampini } 45094f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45104f1b2e48SStefano Zampini } 45114f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 45124f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 45134f1b2e48SStefano Zampini } 45144f1b2e48SStefano Zampini 4515674ae819SStefano Zampini /* Graph's connected components analysis */ 4516674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4517674ae819SStefano Zampini 4518674ae819SStefano Zampini /* print some info to stdout */ 4519674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4520302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4521674ae819SStefano Zampini } 4522674ae819SStefano Zampini PetscFunctionReturn(0); 4523674ae819SStefano Zampini } 4524674ae819SStefano Zampini 4525dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4526674ae819SStefano Zampini #undef __FUNCT__ 4527674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4528dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4529674ae819SStefano Zampini { 4530dc456d91SStefano Zampini PetscSF sf; 4531dc456d91SStefano Zampini PetscLayout map; 4532dc456d91SStefano Zampini const PetscInt *idxs; 4533dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4534dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4535dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4536dc456d91SStefano Zampini PetscMPIInt commsize; 4537674ae819SStefano Zampini PetscBool first_found; 4538674ae819SStefano Zampini PetscErrorCode ierr; 4539674ae819SStefano Zampini 4540674ae819SStefano Zampini PetscFunctionBegin; 4541dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4542dc456d91SStefano Zampini if (subset_mult) { 4543dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4544dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4545dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4546674ae819SStefano Zampini } 4547dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4548dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4549dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4550dc456d91SStefano Zampini for (i=0;i<n;i++) { 4551dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4552dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4553674ae819SStefano Zampini } 4554dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4555b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4556dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4557dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4558dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4559dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4560dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4561dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4562dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4563dc456d91SStefano Zampini 4564dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4565dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4566dc456d91SStefano Zampini if (subset_mult) { 4567dc456d91SStefano Zampini const PetscInt* idxs_mult; 4568dc456d91SStefano Zampini 4569dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4570dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4571dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4572674ae819SStefano Zampini } else { 4573dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4574674ae819SStefano Zampini } 4575dc456d91SStefano Zampini /* local size of new subset */ 4576dc456d91SStefano Zampini n_n = 0; 4577dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4578dc456d91SStefano Zampini 4579dc456d91SStefano Zampini /* global indexes in layout */ 4580dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4581dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4582dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4583dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4584dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4585dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4586dc456d91SStefano Zampini 4587dc456d91SStefano Zampini /* reduce from leaves to roots */ 4588dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 458964a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 459064a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4591dc456d91SStefano Zampini 4592dc456d91SStefano Zampini /* count indexes in local part of layout */ 4593674ae819SStefano Zampini nlocals = 0; 4594674ae819SStefano Zampini first_index = -1; 4595674ae819SStefano Zampini first_found = PETSC_FALSE; 4596dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4597dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4598674ae819SStefano Zampini first_found = PETSC_TRUE; 4599674ae819SStefano Zampini first_index = i; 4600674ae819SStefano Zampini } 4601dc456d91SStefano Zampini nlocals += root_data[i]; 4602674ae819SStefano Zampini } 4603dc456d91SStefano Zampini 4604dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 46055fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4606dc456d91SStefano Zampini start = 0; 460764a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46085fa240b1SStefano Zampini #else 460964a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46105fa240b1SStefano Zampini start = start-nlocals; 46115fa240b1SStefano Zampini #endif 46125fa240b1SStefano Zampini 4613dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4614dc456d91SStefano Zampini *N_n = start + nlocals; 4615dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4616dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4617674ae819SStefano Zampini } 46185fa240b1SStefano Zampini 46195fa240b1SStefano Zampini /* adapt root data with cumulative */ 4620674ae819SStefano Zampini if (first_found) { 4621dc456d91SStefano Zampini PetscInt old_index; 4622dc456d91SStefano Zampini 4623dc456d91SStefano Zampini root_data[first_index] += start; 4624674ae819SStefano Zampini old_index = first_index; 4625dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4626dc456d91SStefano Zampini if (root_data[i]) { 4627dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4628674ae819SStefano Zampini old_index = i; 4629674ae819SStefano Zampini } 4630674ae819SStefano Zampini } 4631674ae819SStefano Zampini } 4632dc456d91SStefano Zampini 4633dc456d91SStefano Zampini /* from roots to leaves */ 4634dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4635dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4636dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4637dc456d91SStefano Zampini 4638dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4639dc456d91SStefano Zampini if (subset_mult) { 4640dc456d91SStefano Zampini const PetscInt* idxs_mult; 4641dc456d91SStefano Zampini PetscInt cum; 4642dc456d91SStefano Zampini 4643dc456d91SStefano Zampini cum = 0; 4644dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4645dc456d91SStefano Zampini for (i=0;i<n;i++) { 4646dc456d91SStefano Zampini PetscInt j; 4647dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4648674ae819SStefano Zampini } 4649dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4650674ae819SStefano Zampini } else { 4651dc456d91SStefano Zampini for (i=0;i<n;i++) { 4652dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4653674ae819SStefano Zampini } 4654674ae819SStefano Zampini } 4655dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4656dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4657674ae819SStefano Zampini PetscFunctionReturn(0); 4658674ae819SStefano Zampini } 46599a7d3425SStefano Zampini 46609a7d3425SStefano Zampini #undef __FUNCT__ 46619a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 46629a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 46639a7d3425SStefano Zampini { 46649a7d3425SStefano Zampini PetscInt i,j; 46659a7d3425SStefano Zampini PetscScalar *alphas; 46669a7d3425SStefano Zampini PetscErrorCode ierr; 46679a7d3425SStefano Zampini 46689a7d3425SStefano Zampini PetscFunctionBegin; 46699a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4670785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 46719a7d3425SStefano Zampini for (i=0;i<n;i++) { 46729a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 46739a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 46749a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 46759a7d3425SStefano Zampini } 46769a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 46779a7d3425SStefano Zampini PetscFunctionReturn(0); 46789a7d3425SStefano Zampini } 46799a7d3425SStefano Zampini 4680e7931f94SStefano Zampini #undef __FUNCT__ 468170cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4682b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4683e7931f94SStefano Zampini { 468452e5ac9dSStefano Zampini IS ranks_send_to; 4685e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4686e7931f94SStefano Zampini PetscMPIInt size,rank,color; 468752e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 468852e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 46893837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 46902b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4691e7931f94SStefano Zampini PetscSubcomm subcomm; 469252e5ac9dSStefano Zampini PetscErrorCode ierr; 4693a57a6d2fSStefano Zampini 4694e7931f94SStefano Zampini PetscFunctionBegin; 4695c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 4696c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4697c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4698e7931f94SStefano Zampini 4699e7931f94SStefano Zampini /* Get info on mapping */ 47003bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 47013bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4702e7931f94SStefano Zampini 4703e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4704785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4705e7931f94SStefano Zampini xadj[0] = 0; 4706e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4707785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4708785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4709e7931f94SStefano Zampini 47102b510759SStefano Zampini if (threshold) { 4711d023bfaeSStefano Zampini PetscInt xadj_count = 0; 47122b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4713d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4714d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4715d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4716d023bfaeSStefano Zampini xadj_count++; 4717e7931f94SStefano Zampini } 4718e7931f94SStefano Zampini } 4719d023bfaeSStefano Zampini xadj[1] = xadj_count; 4720c8587f34SStefano Zampini } else { 4721e7931f94SStefano Zampini if (xadj[1]) { 4722e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4723e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4724c8587f34SStefano Zampini } 4725e7931f94SStefano Zampini } 47263bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4727e7931f94SStefano Zampini if (use_square) { 4728e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4729e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4730e7931f94SStefano Zampini } 4731e7931f94SStefano Zampini } 4732e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4733e7931f94SStefano Zampini 47343837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4735e7931f94SStefano Zampini 4736e7931f94SStefano Zampini /* 4737e7931f94SStefano Zampini Restrict work on active processes only. 4738e7931f94SStefano Zampini */ 4739e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4740e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4741e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 47422b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4743d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4744e7931f94SStefano Zampini if (color) { 4745e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4746e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4747e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4748c8587f34SStefano Zampini } else { 474952e5ac9dSStefano Zampini Mat subdomain_adj; 475052e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 475152e5ac9dSStefano Zampini MatPartitioning partitioner; 475252e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 475352e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4754b0c7d250SStefano Zampini PetscBool aggregate; 4755b0c7d250SStefano Zampini 4756306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4757785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4758e7931f94SStefano Zampini prank = rank; 4759306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 47608002ef2cSStefano Zampini /* 4761e7931f94SStefano Zampini for (i=0;i<size;i++) { 4762e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4763c8587f34SStefano Zampini } 47648002ef2cSStefano Zampini */ 4765e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4766e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4767c8587f34SStefano Zampini } 4768e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4769b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4770b0c7d250SStefano Zampini if (aggregate) { 4771b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4772b0c7d250SStefano Zampini PetscMPIInt nrank; 4773b0c7d250SStefano Zampini PetscScalar *vals; 4774b0c7d250SStefano Zampini 4775b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4776b0c7d250SStefano Zampini lrows = 0; 4777b0c7d250SStefano Zampini if (nrank<redprocs) { 4778b0c7d250SStefano Zampini lrows = size/redprocs; 4779b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4780b0c7d250SStefano Zampini } 47815fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4782b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4783b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4784b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4785b0c7d250SStefano Zampini row = nrank; 4786b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4787b0c7d250SStefano Zampini cols = adjncy; 4788b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4789b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4790b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4791b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4792b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 479352e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 479452e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 479552e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4796b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4797b0c7d250SStefano Zampini } else { 4798306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4799b0c7d250SStefano Zampini } 480022b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4801e7931f94SStefano Zampini 4802e7931f94SStefano Zampini /* Partition */ 4803306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4804e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4805e7931f94SStefano Zampini if (use_vwgt) { 48063837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4807e7931f94SStefano Zampini v_wgt[0] = local_size; 4808e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4809c8587f34SStefano Zampini } 481028143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 481128143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4812e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4813e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 481422b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4815e7931f94SStefano Zampini 481652e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 481752e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 481852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 481952e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4820b0c7d250SStefano Zampini if (!redprocs) { 4821b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 482228143c3dSStefano Zampini } else { 4823b0c7d250SStefano Zampini PetscInt idxs[1]; 4824b0c7d250SStefano Zampini PetscMPIInt tag; 4825b0c7d250SStefano Zampini MPI_Request *reqs; 4826b0c7d250SStefano Zampini 4827b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4828b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4829b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4830b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 483128143c3dSStefano Zampini } 4832b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4833b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4834b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4835b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4836e7931f94SStefano Zampini } 483752e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4838e7931f94SStefano Zampini /* clean up */ 4839e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 484052e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4841e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4842e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4843e7931f94SStefano Zampini } 4844e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4845e7931f94SStefano Zampini 4846e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4847e7931f94SStefano Zampini i = 1; 4848e7931f94SStefano Zampini if (color) i=0; 4849e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4850e7931f94SStefano Zampini /* get back IS */ 4851e7931f94SStefano Zampini *is_sends = ranks_send_to; 4852e7931f94SStefano Zampini PetscFunctionReturn(0); 4853e7931f94SStefano Zampini } 4854e7931f94SStefano Zampini 4855e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4856e7931f94SStefano Zampini 4857e7931f94SStefano Zampini #undef __FUNCT__ 4858e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 485953a05cb3SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 4860e7931f94SStefano Zampini { 486170cf5478SStefano Zampini Mat local_mat; 4862e7931f94SStefano Zampini IS is_sends_internal; 48639d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 486428143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 48659d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4866e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4867e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4868e7931f94SStefano Zampini const PetscInt* is_indices; 4869e7931f94SStefano Zampini MatType new_local_type; 4870e7931f94SStefano Zampini /* buffers */ 4871e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 487228143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 48739d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4874e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4875e7931f94SStefano Zampini /* MPI */ 487628143c3dSStefano Zampini MPI_Comm comm,comm_n; 487728143c3dSStefano Zampini PetscSubcomm subcomm; 4878e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 487928143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 488028143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 488128143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 488228143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 488328143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4884e7931f94SStefano Zampini PetscErrorCode ierr; 4885e7931f94SStefano Zampini 4886e7931f94SStefano Zampini PetscFunctionBegin; 488728143c3dSStefano Zampini /* TODO: add missing checks */ 488828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 488928143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 489028143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 489128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4892e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 489328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4894e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4895e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4896e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4897e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4898e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 489928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 490070cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 490170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 490228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 490370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 490470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 490570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 490670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 490770cf5478SStefano Zampini } 4908e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4909e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4910e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4911e7931f94SStefano Zampini if (!is_sends) { 491228143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4913b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4914c8587f34SStefano Zampini } else { 4915e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4916e7931f94SStefano Zampini is_sends_internal = is_sends; 4917c8587f34SStefano Zampini } 4918e7931f94SStefano Zampini 4919e7931f94SStefano Zampini /* get comm */ 4920a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4921e7931f94SStefano Zampini 4922e7931f94SStefano Zampini /* compute number of sends */ 4923e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4924e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4925e7931f94SStefano Zampini 4926e7931f94SStefano Zampini /* compute number of receives */ 4927e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4928785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4929e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4930e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4931e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4932e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4933e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4934e7931f94SStefano Zampini 493528143c3dSStefano Zampini /* restrict comm if requested */ 493628143c3dSStefano Zampini subcomm = 0; 493728143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 493828143c3dSStefano Zampini if (restrict_comm) { 4939779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4940779c1cceSStefano Zampini 494128143c3dSStefano Zampini color = 0; 494253a05cb3SStefano Zampini if (restrict_full) { 494353a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 494453a05cb3SStefano Zampini } else { 494553a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 494653a05cb3SStefano Zampini } 4947b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 494828143c3dSStefano Zampini subcommsize = commsize - subcommsize; 494928143c3dSStefano Zampini /* check if reuse has been requested */ 495028143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 495128143c3dSStefano Zampini if (*mat_n) { 495228143c3dSStefano Zampini PetscMPIInt subcommsize2; 495328143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 495428143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 495528143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 495628143c3dSStefano Zampini } else { 495728143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 495828143c3dSStefano Zampini } 495928143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4960779c1cceSStefano Zampini PetscMPIInt rank; 4961779c1cceSStefano Zampini 4962779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 496328143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 496428143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 496528143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4966306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 496728143c3dSStefano Zampini } 496828143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 496928143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 497028143c3dSStefano Zampini } else { 497128143c3dSStefano Zampini comm_n = comm; 497228143c3dSStefano Zampini } 497328143c3dSStefano Zampini 4974e7931f94SStefano Zampini /* prepare send/receive buffers */ 4975785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4976e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4977785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4978e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 497928143c3dSStefano Zampini if (nis) { 4980854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 498128143c3dSStefano Zampini } 4982e7931f94SStefano Zampini 498328143c3dSStefano Zampini /* Get data from local matrices */ 49846c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4985e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4986e7931f94SStefano Zampini /* 4987e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4988e7931f94SStefano Zampini send_buffer_idxs should contain: 4989e7931f94SStefano Zampini - MatType_PRIVATE type 4990e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4991e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4992e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4993e7931f94SStefano Zampini */ 49946c4ed002SBarry Smith else { 4995e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 49963bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4997854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4998e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4999e7931f94SStefano Zampini send_buffer_idxs[1] = i; 50003bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5001e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 50023bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5003e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5004e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5005e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5006e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5007c8587f34SStefano Zampini } 5008c8587f34SStefano Zampini } 5009e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 501028143c3dSStefano Zampini /* additional is (if any) */ 501128143c3dSStefano Zampini if (nis) { 501228143c3dSStefano Zampini PetscMPIInt psum; 501328143c3dSStefano Zampini PetscInt j; 501428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 501528143c3dSStefano Zampini PetscInt plen; 501628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 501728143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 501828143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 501928143c3dSStefano Zampini } 5020854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 502128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 502228143c3dSStefano Zampini PetscInt plen; 502328143c3dSStefano Zampini const PetscInt *is_array_idxs; 502428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 502528143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 502628143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 502728143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 502828143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 502928143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 503028143c3dSStefano Zampini } 503128143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 503228143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 503328143c3dSStefano Zampini } 503428143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 503528143c3dSStefano Zampini } 503628143c3dSStefano Zampini 5037e7931f94SStefano Zampini buf_size_idxs = 0; 5038e7931f94SStefano Zampini buf_size_vals = 0; 503928143c3dSStefano Zampini buf_size_idxs_is = 0; 5040e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5041e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5042e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 504328143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5044e7931f94SStefano Zampini } 5045785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5046785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 504795ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5048e7931f94SStefano Zampini 5049e7931f94SStefano Zampini /* get new tags for clean communications */ 5050e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5051e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 505228143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5053e7931f94SStefano Zampini 5054e7931f94SStefano Zampini /* allocate for requests */ 5055785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5056785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 505795ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5058785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5059785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 506095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5061e7931f94SStefano Zampini 5062e7931f94SStefano Zampini /* communications */ 5063e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5064e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 506528143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5066e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5067e7931f94SStefano Zampini source_dest = onodes[i]; 5068e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5069e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5070e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5071e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 507228143c3dSStefano Zampini if (nis) { 507328143c3dSStefano 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); 507428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 507528143c3dSStefano Zampini } 5076e7931f94SStefano Zampini } 5077e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5078e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5079e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5080e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 508128143c3dSStefano Zampini if (nis) { 508228143c3dSStefano 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); 508328143c3dSStefano Zampini } 5084e7931f94SStefano Zampini } 5085e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5086e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5087e7931f94SStefano Zampini 5088e7931f94SStefano Zampini /* assemble new l2g map */ 5089e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5090e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50919d30be91SStefano Zampini new_local_rows = 0; 5092e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50939d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5094e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5095e7931f94SStefano Zampini } 50969d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5097e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50989d30be91SStefano Zampini new_local_rows = 0; 5099e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51009d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 51019d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5102e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5103e7931f94SStefano Zampini } 51049d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 51059d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5106e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5107e7931f94SStefano Zampini 5108e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5109e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5110e7931f94SStefano 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) */ 5111e7931f94SStefano Zampini if (n_recvs) { 511228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5113e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5114e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5115e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5116e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5117e7931f94SStefano Zampini break; 5118e7931f94SStefano Zampini } 5119e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5120e7931f94SStefano Zampini } 5121e7931f94SStefano Zampini switch (new_local_type_private) { 512228143c3dSStefano Zampini case MATDENSE_PRIVATE: 512328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5124e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5125e7931f94SStefano Zampini bs = 1; 512628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 512728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 512828143c3dSStefano Zampini bs = 1; 512928143c3dSStefano Zampini } 5130e7931f94SStefano Zampini break; 5131e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5132e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5133e7931f94SStefano Zampini bs = 1; 5134e7931f94SStefano Zampini break; 5135e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5136e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5137e7931f94SStefano Zampini break; 5138e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5139e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5140e7931f94SStefano Zampini break; 5141e7931f94SStefano Zampini default: 51429d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5143e7931f94SStefano Zampini break; 5144e7931f94SStefano Zampini } 514528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 514628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 514728143c3dSStefano Zampini bs = 1; 5148e7931f94SStefano Zampini } 5149e7931f94SStefano Zampini 515070cf5478SStefano Zampini /* create MATIS object if needed */ 515170cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5152e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5153e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 515470cf5478SStefano Zampini } else { 515570cf5478SStefano Zampini /* it also destroys the local matrices */ 515670cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 515770cf5478SStefano Zampini } 515870cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5159e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 51609d30be91SStefano Zampini 51619d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 51629d30be91SStefano Zampini 51639d30be91SStefano Zampini /* Global to local map of received indices */ 51649d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 51659d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 51669d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 51679d30be91SStefano Zampini 51689d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 51699d30be91SStefano Zampini buf_size_idxs = 0; 51709d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51719d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 51729d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 51739d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 51749d30be91SStefano Zampini } 51759d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 51769d30be91SStefano Zampini 51779d30be91SStefano Zampini /* set preallocation */ 51789d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 51799d30be91SStefano Zampini if (!newisdense) { 51809d30be91SStefano Zampini PetscInt *new_local_nnz=0; 51819d30be91SStefano Zampini 51829d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 51839d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 51849d30be91SStefano Zampini if (n_recvs) { 51859d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 51869d30be91SStefano Zampini } 51879d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51889d30be91SStefano Zampini PetscInt j; 51899d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 51909d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 51919d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 51929d30be91SStefano Zampini } 51939d30be91SStefano Zampini } else { 51949d30be91SStefano Zampini /* TODO */ 51959d30be91SStefano Zampini } 51969d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 51979d30be91SStefano Zampini } 51989d30be91SStefano Zampini if (new_local_nnz) { 51999d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 52009d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 52019d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 52029d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52039d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 52049d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52059d30be91SStefano Zampini } else { 52069d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52079d30be91SStefano Zampini } 52089d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 52099d30be91SStefano Zampini } else { 52109d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52119d30be91SStefano Zampini } 5212e7931f94SStefano Zampini 5213e7931f94SStefano Zampini /* set values */ 5214e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 52159d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5216e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5217e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5218e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 52199d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5220e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5221e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5222e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 522328143c3dSStefano Zampini } else { 522428143c3dSStefano Zampini /* TODO */ 5225e7931f94SStefano Zampini } 5226e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5227e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5228e7931f94SStefano Zampini } 5229e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5230e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 523170cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 523270cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 52339d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 52349d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5235e7931f94SStefano Zampini 5236dfd14d43SStefano Zampini #if 0 523728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5238e7931f94SStefano Zampini Vec lvec,rvec; 5239e7931f94SStefano Zampini PetscReal infty_error; 5240e7931f94SStefano Zampini 52412a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5242e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5243e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5244e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 524570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5246e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5247e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5248e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5249e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5250e7931f94SStefano Zampini } 525128143c3dSStefano Zampini #endif 5252e7931f94SStefano Zampini 525328143c3dSStefano Zampini /* assemble new additional is (if any) */ 525428143c3dSStefano Zampini if (nis) { 525528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 525628143c3dSStefano Zampini 525728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5258854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 525928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 526028143c3dSStefano Zampini psum = 0; 526128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 526228143c3dSStefano Zampini for (j=0;j<nis;j++) { 526328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 526428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 526528143c3dSStefano Zampini psum += plen; 526628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 526728143c3dSStefano Zampini } 526828143c3dSStefano Zampini } 5269854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5270854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 527128143c3dSStefano Zampini for (i=1;i<nis;i++) { 527228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 527328143c3dSStefano Zampini } 527428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 527528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 527628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 527728143c3dSStefano Zampini for (j=0;j<nis;j++) { 527828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 527928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 528028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 528128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 528228143c3dSStefano Zampini } 528328143c3dSStefano Zampini } 528428143c3dSStefano Zampini for (i=0;i<nis;i++) { 528528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 528628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 528728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 528828143c3dSStefano Zampini } 528928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 529028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 529128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 529228143c3dSStefano Zampini } 5293e7931f94SStefano Zampini /* free workspace */ 529428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5295e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5296e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5297e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5298e7931f94SStefano Zampini if (isdense) { 5299e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5300e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5301e7931f94SStefano Zampini } else { 5302e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5303e7931f94SStefano Zampini } 530428143c3dSStefano Zampini if (nis) { 530528143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 530628143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 530728143c3dSStefano Zampini } 5308e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5309e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 531028143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5311e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5312e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 531328143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5314e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5315e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5316e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5317e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5318e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 531928143c3dSStefano Zampini if (nis) { 532028143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 532128143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 532228143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 532328143c3dSStefano Zampini } 532428143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 532528143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 532628143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 532728143c3dSStefano Zampini for (i=0;i<nis;i++) { 532828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 532928143c3dSStefano Zampini } 533053a05cb3SStefano Zampini *mat_n = NULL; 533128143c3dSStefano Zampini } 5332e7931f94SStefano Zampini PetscFunctionReturn(0); 5333e7931f94SStefano Zampini } 5334a57a6d2fSStefano Zampini 533512edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5336af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 533712edc857SStefano Zampini 5338c8587f34SStefano Zampini #undef __FUNCT__ 5339c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5340c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5341c8587f34SStefano Zampini { 5342c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5343c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 534420a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 53459881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 534620a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 53476e683305SStefano Zampini IS coarse_is,*isarray; 53486e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 534930368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5350f9eb5b7dSStefano Zampini PC pc_temp; 5351c8587f34SStefano Zampini PCType coarse_pc_type; 5352c8587f34SStefano Zampini KSPType coarse_ksp_type; 5353f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 53544f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 53556e683305SStefano Zampini Mat t_coarse_mat_is; 53566e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 53576e683305SStefano Zampini PetscMPIInt all_procs; 535874e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 535968457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 536022bc73bbSStefano Zampini PetscScalar *array; 53619881197aSStefano Zampini PetscErrorCode ierr; 5362fdc09c96SStefano Zampini 5363c8587f34SStefano Zampini PetscFunctionBegin; 5364c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 536568457ee5SStefano 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 */ 5366fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 53675a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5368fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5369f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5370f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5371f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5372fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 537351bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 537451bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5375dc4bcba2SStefano Zampini PC pc; 5376dc4bcba2SStefano Zampini PetscBool isbddc; 5377dc4bcba2SStefano Zampini 5378dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5379dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5380dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5381dc4bcba2SStefano Zampini if (isbddc) { 538263c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 538363c961adSStefano Zampini } else { 5384727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 538563c961adSStefano Zampini } 5386fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5387fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5388fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5389f4ddd8eeSStefano Zampini } 5390fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5391fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5392f4ddd8eeSStefano Zampini } 539370cf5478SStefano Zampini /* reset any subassembling information */ 539470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 53956e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 53966e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5397fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5398f4ddd8eeSStefano Zampini } 5399c8587f34SStefano Zampini 54006e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 54012b510759SStefano Zampini im_active = !!(pcis->n); 5402b2566f29SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 54036e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 54046e683305SStefano Zampini void_procs = all_procs-active_procs; 54056e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 540674e2c79eSStefano Zampini redist = PETSC_FALSE; 540722bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 54086e683305SStefano Zampini csin_ml = PETSC_TRUE; 54096e683305SStefano Zampini ncoarse_ml = void_procs; 5410779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5411779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 54126e683305SStefano Zampini csin_ds = PETSC_TRUE; 541318a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 541418a45a71SStefano Zampini redist = PETSC_TRUE; 541518a45a71SStefano Zampini } else { 54166e683305SStefano Zampini csin_ds = PETSC_TRUE; 5417779c1cceSStefano Zampini ncoarse_ds = active_procs; 5418779c1cceSStefano Zampini redist = PETSC_TRUE; 541918a45a71SStefano Zampini } 54206e683305SStefano Zampini } else { 54216e683305SStefano Zampini csin_ml = PETSC_FALSE; 54226e683305SStefano Zampini ncoarse_ml = all_procs; 54236e683305SStefano Zampini if (void_procs) { 54246e683305SStefano Zampini csin_ds = PETSC_TRUE; 54256e683305SStefano Zampini ncoarse_ds = void_procs; 54266e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 54276e683305SStefano Zampini } else { 5428779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 542974e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 543074e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 543174e2c79eSStefano Zampini redist = PETSC_TRUE; 543274e2c79eSStefano Zampini } else { 54336e683305SStefano Zampini csin_ds = PETSC_FALSE; 54346e683305SStefano Zampini ncoarse_ds = all_procs; 54356e683305SStefano Zampini } 54366e683305SStefano Zampini } 543774e2c79eSStefano Zampini } 54386e683305SStefano Zampini 54396e683305SStefano Zampini /* 54406e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 54416e683305SStefano Zampini - we have not exceeded the number of levels requested 54426e683305SStefano Zampini - we can actually subassemble the active processes 54436e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 54446e683305SStefano Zampini */ 54456e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 54466e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 54476e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 54486e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 54496e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5450f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 54512b510759SStefano Zampini } else { 5452f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5453c8587f34SStefano Zampini } 5454c8587f34SStefano Zampini } 54556e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 54566e683305SStefano Zampini if (multilevel_allowed) { 54576e683305SStefano Zampini ncoarse = ncoarse_ml; 54586e683305SStefano Zampini csin = csin_ml; 545958da7f69SStefano Zampini redist = PETSC_FALSE; 54606e683305SStefano Zampini } else { 54616e683305SStefano Zampini ncoarse = ncoarse_ds; 54626e683305SStefano Zampini csin = csin_ds; 54636e683305SStefano Zampini } 5464e7931f94SStefano Zampini 5465abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5466abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5467abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5468abbbba34SStefano Zampini 5469abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 547022bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 547122bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 547222bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 547322bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5474e176bc59SStefano 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); 54756e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 54766e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 54776e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5478abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5479abbbba34SStefano Zampini 54806e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 548130368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 54826e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 54836e683305SStefano Zampini const PetscInt *idxs; 54846e683305SStefano Zampini ISLocalToGlobalMapping tmap; 54856e683305SStefano Zampini 54866e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 54870be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 54886e683305SStefano Zampini /* allocate space for temporary storage */ 5489854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5490854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 54916e683305SStefano Zampini /* allocate for IS array */ 54926e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 54936e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 549430368db7SStefano Zampini nisvert = 0; 549530368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 549630368db7SStefano Zampini nisvert = 1; 549730368db7SStefano Zampini } 549830368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5499854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 55006e683305SStefano Zampini /* dofs splitting */ 55016e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 55026e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 55036e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 55046e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55056e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55066e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55076e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 550830368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 55096e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 55106e683305SStefano Zampini } 55116e683305SStefano Zampini /* neumann boundaries */ 55126e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 55136e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 55146e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 55156e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55166e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55176e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55186e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 551930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 55206e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 55216e683305SStefano Zampini } 552230368db7SStefano Zampini /* primal vertices (benign) */ 552330368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 552430368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 552530368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 552630368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 552730368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 552830368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 552930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 553030368db7SStefano Zampini } 55316e683305SStefano Zampini /* free memory */ 55326e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 55336e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 55346e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 55356e683305SStefano Zampini } else { 55366e683305SStefano Zampini nis = 0; 55376e683305SStefano Zampini nisdofs = 0; 55386e683305SStefano Zampini nisneu = 0; 553930368db7SStefano Zampini nisvert = 0; 55406e683305SStefano Zampini isarray = NULL; 55416e683305SStefano Zampini } 55426e683305SStefano Zampini /* destroy no longer needed map */ 55436e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 55446e683305SStefano Zampini 55456e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 55466e683305SStefano Zampini coarse_mat_is = NULL; 55476e683305SStefano Zampini if (csin) { 55486e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 554974e2c79eSStefano Zampini if (redist) { 555074e2c79eSStefano Zampini PetscMPIInt rank; 5551779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 555274e2c79eSStefano Zampini 555374e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 555458da7f69SStefano Zampini spc = active_procs/ncoarse; 555558da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5556779c1cceSStefano Zampini if (im_active) { 5557779c1cceSStefano Zampini destsize = 1; 555874e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 555974e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 556074e2c79eSStefano Zampini } else { 556174e2c79eSStefano Zampini dest[0] = rank/(spc+1); 556274e2c79eSStefano Zampini } 556374e2c79eSStefano Zampini } else { 5564779c1cceSStefano Zampini destsize = 0; 55656e683305SStefano Zampini } 5566779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5567779c1cceSStefano Zampini } else if (csin_type_simple) { 55686e683305SStefano Zampini PetscMPIInt rank; 55696e683305SStefano Zampini PetscInt issize,isidx; 5570779c1cceSStefano Zampini 55716e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 55726e683305SStefano Zampini if (im_active) { 55736e683305SStefano Zampini issize = 1; 55746e683305SStefano Zampini isidx = (PetscInt)rank; 55756e683305SStefano Zampini } else { 55766e683305SStefano Zampini issize = 0; 55776e683305SStefano Zampini isidx = -1; 55786e683305SStefano Zampini } 55796e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5580779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5581b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 55826e683305SStefano Zampini } 5583779c1cceSStefano Zampini 5584779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5585779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5586779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5587779c1cceSStefano Zampini PetscInt *coarse_candidates; 5588779c1cceSStefano Zampini const PetscInt* tisindices; 5589779c1cceSStefano Zampini 5590779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5591779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5592779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5593779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5594779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5595779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5596779c1cceSStefano Zampini } 5597779c1cceSStefano Zampini } 5598779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5599779c1cceSStefano Zampini 5600779c1cceSStefano Zampini 56016e683305SStefano Zampini if (pcbddc->dbg_flag) { 56026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 56046e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5606779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 56076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 56086e683305SStefano Zampini } 56096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 56106e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56116e683305SStefano Zampini } 56126e683305SStefano Zampini /* shift the pattern on coarse candidates */ 56136e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 56146e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5615854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 56166e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 56176e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 56186e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 56196e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 56206e683305SStefano Zampini } 56216e683305SStefano Zampini if (pcbddc->dbg_flag) { 56226e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 56246e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56256e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56266e683305SStefano Zampini } 5627779c1cceSStefano Zampini } 56286e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 562953a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 563053a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_FALSE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 563153a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 563253a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 563353a05cb3SStefano Zampini } 56346e683305SStefano Zampini } else { 56356e683305SStefano Zampini if (pcbddc->dbg_flag) { 56366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56376e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 56386e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56396e683305SStefano Zampini } 56406e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 56416e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 56426e683305SStefano Zampini } 56436e683305SStefano Zampini 56446e683305SStefano Zampini /* create local to global scatters for coarse problem */ 564568457ee5SStefano Zampini if (compute_vecs) { 56466e683305SStefano Zampini PetscInt lrows; 56476e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 56486e683305SStefano Zampini if (coarse_mat_is) { 56496e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 56506e683305SStefano Zampini } else { 56516e683305SStefano Zampini lrows = 0; 56526e683305SStefano Zampini } 56536e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 56546e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 56556e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 56566e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56576e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56586e683305SStefano Zampini } 56596e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 56606e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5661c8587f34SStefano Zampini 5662f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5663f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5664f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5665f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5666f9eb5b7dSStefano Zampini } else { 5667f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5668f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5669c8587f34SStefano Zampini } 5670c8587f34SStefano Zampini 56716e683305SStefano Zampini /* print some info if requested */ 56726e683305SStefano Zampini if (pcbddc->dbg_flag) { 56736e683305SStefano Zampini if (!multilevel_allowed) { 56746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56756e683305SStefano Zampini if (multilevel_requested) { 56766e683305SStefano 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); 56776e683305SStefano Zampini } else if (pcbddc->max_levels) { 56786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 56796e683305SStefano Zampini } 56806e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56816e683305SStefano Zampini } 56826e683305SStefano Zampini } 56836e683305SStefano Zampini 5684f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 56856e683305SStefano Zampini if (coarse_mat_is) { 56866e683305SStefano Zampini MatReuse coarse_mat_reuse; 56876a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 56886e683305SStefano Zampini if (pcbddc->dbg_flag) { 56896e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 56906e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56916e683305SStefano Zampini } 5692f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5693312be037SStefano Zampini char prefix[256],str_level[16]; 5694e604994aSStefano Zampini size_t len; 56956e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5696422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5697c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5698f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 56995f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5700c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 57016e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5702c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5703c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5704e604994aSStefano Zampini /* prefix */ 5705e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5706e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5707e604994aSStefano Zampini if (!pcbddc->current_level) { 5708e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5709e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5710c8587f34SStefano Zampini } else { 5711e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5712312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5713312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 571434d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5715312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5716e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5717e604994aSStefano Zampini } 5718e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 57193e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 57203e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 57213e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 57223e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5723f9eb5b7dSStefano Zampini /* allow user customization */ 5724f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 57253e3c6dadSStefano Zampini } 57263e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 572751bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 57283e3c6dadSStefano Zampini if (nisdofs) { 57293e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 57303e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 57313e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 57323e3c6dadSStefano Zampini } 57333e3c6dadSStefano Zampini } 57343e3c6dadSStefano Zampini if (nisneu) { 57353e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 57363e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5737312be037SStefano Zampini } 573830368db7SStefano Zampini if (nisvert) { 573930368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 574030368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 574130368db7SStefano Zampini } 5742f9eb5b7dSStefano Zampini 5743f9eb5b7dSStefano Zampini /* get some info after set from options */ 5744f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5745f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 57464f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 57476e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5748f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5749f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5750f9eb5b7dSStefano Zampini } 575139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 57524f3a063dSStefano Zampini if (isredundant) { 57534f3a063dSStefano Zampini KSP inner_ksp; 57544f3a063dSStefano Zampini PC inner_pc; 57554f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 57564f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 57574f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 57584f3a063dSStefano Zampini } 5759f9eb5b7dSStefano Zampini 5760f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5761fa7f1dd8SStefano Zampini if (coarse_reuse) { 576281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5763fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 57646e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5765fa7f1dd8SStefano Zampini } else { 57666e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5767fa7f1dd8SStefano Zampini } 5768c8587f34SStefano Zampini if (isbddc || isnn) { 5769720d30f9SStefano Zampini if (isbddc) { /* currently there are no APIs for these options */ 5770720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5771720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 5772720d30f9SStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_saddle_point; 5773720d30f9SStefano Zampini pcbddc_coarse->benign_compute_nonetflux = pcbddc->benign_compute_nonetflux; 5774720d30f9SStefano Zampini if (pcbddc_coarse->benign_compute_nonetflux) { 5775720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 5776720d30f9SStefano Zampini } 5777d4d8cf7bSStefano Zampini } 577822bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 577970cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5780b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 578122b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 57826e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 57846e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 57856e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 578622b6e8a2SStefano Zampini } 578770cf5478SStefano Zampini } 578853a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 578970cf5478SStefano Zampini } else { 579022bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 579122bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 579222bc73bbSStefano Zampini } 579322bc73bbSStefano Zampini } else { 57942e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5795c8587f34SStefano Zampini } 5796c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5797c8587f34SStefano Zampini 57983301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 57995a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 58003301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 58013301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 58023301b35fSStefano Zampini } 58033301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 58043301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 58053301b35fSStefano Zampini } 58063301b35fSStefano Zampini if (pc->pmat->spd_set) { 58073301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 58083301b35fSStefano Zampini } 58096e683305SStefano Zampini /* set operators */ 58105f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 58116e683305SStefano Zampini if (pcbddc->dbg_flag) { 58126e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 58136e683305SStefano Zampini } 58146e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 58156e683305SStefano Zampini coarse_mat = 0; 58166e683305SStefano Zampini } 58176e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5818b1ecc7b1SStefano Zampini #if 0 5819b9b85e73SStefano Zampini { 5820b9b85e73SStefano Zampini PetscViewer viewer; 5821b9b85e73SStefano Zampini char filename[256]; 5822b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5823b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 58246a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5825b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5826f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 5827b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5828b9b85e73SStefano Zampini } 5829b9b85e73SStefano Zampini #endif 5830c8587f34SStefano Zampini 5831c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5832298c0119SStefano Zampini #if 0 5833c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5834c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 583598a51de6SStefano Zampini } 5836298c0119SStefano Zampini #endif 5837b0f5fe93SStefano Zampini /* hack */ 583898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 583998a51de6SStefano Zampini Vec crhs,csol; 584004708bb6SStefano Zampini 5841f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5842f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5843f347579bSStefano Zampini if (!csol) { 58442a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5845f9eb5b7dSStefano Zampini } 5846f347579bSStefano Zampini if (!crhs) { 58472a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5848f347579bSStefano Zampini } 5849b0f5fe93SStefano Zampini } 5850b0f5fe93SStefano Zampini 5851b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5852b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5853b0f5fe93SStefano Zampini 5854b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 58554f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58564f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 58574f1b2e48SStefano Zampini } 5858b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5859b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5860b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5861b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5862b0f5fe93SStefano Zampini if (coarse_mat) { 5863b0f5fe93SStefano Zampini Vec nullv; 5864b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5865b0f5fe93SStefano Zampini PetscInt nl; 5866b0f5fe93SStefano Zampini 5867b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5868b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5869b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5870b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5871b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5872b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5873b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5874b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5875b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5876b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5877b0f5fe93SStefano Zampini } 5878b0f5fe93SStefano Zampini } 5879b0f5fe93SStefano Zampini 5880b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5881b0f5fe93SStefano Zampini PetscBool ispreonly; 5882b0f5fe93SStefano Zampini 5883b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5884b0f5fe93SStefano Zampini PetscBool isnull; 5885b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5886d4d8cf7bSStefano Zampini if (0) { 588730368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5888b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5889b0f5fe93SStefano Zampini } else { 5890b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5891b0f5fe93SStefano Zampini } 5892b0f5fe93SStefano Zampini } else { 5893b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5894b0f5fe93SStefano Zampini } 5895b0f5fe93SStefano Zampini } 5896b0f5fe93SStefano Zampini /* setup coarse ksp */ 5897b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5898cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5899cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 59006e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5901c8587f34SStefano Zampini KSP check_ksp; 59022b510759SStefano Zampini KSPType check_ksp_type; 5903c8587f34SStefano Zampini PC check_pc; 59046e683305SStefano Zampini Vec check_vec,coarse_vec; 59056a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 59062b510759SStefano Zampini PetscInt its; 59076e683305SStefano Zampini PetscBool compute_eigs; 59086e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 59096e683305SStefano Zampini PetscInt neigs; 59108e185a42SStefano Zampini const char *prefix; 5911c8587f34SStefano Zampini 59122b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 59136e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5914422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 591523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5916f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 59172b510759SStefano Zampini if (ispreonly) { 59182b510759SStefano Zampini check_ksp_type = KSPPREONLY; 59196e683305SStefano Zampini compute_eigs = PETSC_FALSE; 59202b510759SStefano Zampini } else { 5921cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 59226e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5923c8587f34SStefano Zampini } 5924c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 59256e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 59266e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 59276e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5928a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5929a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5930a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5931a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5932c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5933c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5934c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5935c8587f34SStefano Zampini /* create random vec */ 59366e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 59376e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5938c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5939c8587f34SStefano Zampini if (CoarseNullSpace) { 5940c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5941c8587f34SStefano Zampini } 59426e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5943c8587f34SStefano Zampini /* solve coarse problem */ 59446e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5945c8587f34SStefano Zampini if (CoarseNullSpace) { 59466e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5947c8587f34SStefano Zampini } 5948cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 59496e683305SStefano Zampini if (compute_eigs) { 5950854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5951854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 59526e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 59536e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 59546e683305SStefano Zampini lambda_min = eigs_r[0]; 59556e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 59566e683305SStefano Zampini if (lambda_max>lambda_min) { 5957cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5958cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5959cbcc2c2aSStefano Zampini } 5960c8587f34SStefano Zampini } 5961c8587f34SStefano Zampini } 5962cbcc2c2aSStefano Zampini 5963c8587f34SStefano Zampini /* check coarse problem residual error */ 59646e683305SStefano Zampini if (pcbddc->dbg_flag) { 59656e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 59666e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59676e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5968c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 59696e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 59706e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5971c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5972779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 59736e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 59746e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 59756e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 59766e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5977b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5978b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5979b0f5fe93SStefano Zampini } 59806e683305SStefano Zampini if (compute_eigs) { 59816e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5982deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5983c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 59846e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 59856e683305SStefano 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); 59866e683305SStefano Zampini for (i=0;i<neigs;i++) { 59876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5988c8587f34SStefano Zampini } 59896e683305SStefano Zampini } 59906e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 59916e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59926e683305SStefano Zampini } 5993c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 59946e683305SStefano Zampini if (compute_eigs) { 59956e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 59966e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5997c8587f34SStefano Zampini } 59986e683305SStefano Zampini } 59996e683305SStefano Zampini } 6000cbcc2c2aSStefano Zampini /* print additional info */ 6001cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 60026e683305SStefano Zampini /* waits until all processes reaches this point */ 60036e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6004cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6005cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6006cbcc2c2aSStefano Zampini } 6007cbcc2c2aSStefano Zampini 60082b510759SStefano Zampini /* free memory */ 6009c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6010fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6011c8587f34SStefano Zampini PetscFunctionReturn(0); 6012c8587f34SStefano Zampini } 6013674ae819SStefano Zampini 6014f34684f1SStefano Zampini #undef __FUNCT__ 6015f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6016f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6017f34684f1SStefano Zampini { 6018f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6019f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6020f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6021dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6022dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 602373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6024dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6025f34684f1SStefano Zampini PetscErrorCode ierr; 6026f34684f1SStefano Zampini 6027f34684f1SStefano Zampini PetscFunctionBegin; 6028f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 60296c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6030dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 60313bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6032dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6033dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6034dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6035dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6036dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6037dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60386c4ed002SBarry 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); 6039dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6040dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6041dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6042dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6043dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6044f34684f1SStefano Zampini 6045f34684f1SStefano Zampini /* check numbering */ 6046f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6047019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6048dc456d91SStefano Zampini PetscInt i; 6049b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6050f34684f1SStefano Zampini 6051f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6052f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6053f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 60541575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6055019a44ceSStefano Zampini /* counter */ 6056019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6057019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6058019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6059019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6060019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6061019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6062f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6063f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6064727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6065f34684f1SStefano Zampini } 6066f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6067f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6068f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6069e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6070e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6071e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6072e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6073f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6074019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6075f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6076019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 607775c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 607875c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6079b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6080019a44ceSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a %d processes instead of %d!\n",PetscGlobalRank,i,owned,neigh);CHKERRQ(ierr); 6081f34684f1SStefano Zampini } 6082f34684f1SStefano Zampini } 6083019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6084b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6085f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6086f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6087f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6088f34684f1SStefano Zampini } 6089f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6090f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6091e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6092e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6093f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6094f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6095b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6096ca8b9ea9SStefano Zampini PetscInt *gidxs; 6097ca8b9ea9SStefano Zampini 6098ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 60993bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6100f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6101f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6102f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6103f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 61044bc2dc4bSStefano 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); 6105f34684f1SStefano Zampini } 6106f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6107ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6108f34684f1SStefano Zampini } 6109f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 61101575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6111302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6112f34684f1SStefano Zampini } 61138bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6114f34684f1SStefano Zampini /* get back data */ 6115f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6116f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6117674ae819SStefano Zampini PetscFunctionReturn(0); 6118674ae819SStefano Zampini } 6119674ae819SStefano Zampini 6120e456f2a8SStefano Zampini #undef __FUNCT__ 6121e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6122a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6123e456f2a8SStefano Zampini { 6124e456f2a8SStefano Zampini IS localis_t; 6125a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6126e456f2a8SStefano Zampini PetscScalar *vals; 6127e456f2a8SStefano Zampini PetscErrorCode ierr; 6128e456f2a8SStefano Zampini 6129e456f2a8SStefano Zampini PetscFunctionBegin; 6130a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6131e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6132854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6133e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6134e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6135a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6136a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61371035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6138a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61391035eff8SStefano Zampini } 6140a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6141e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6142e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6143a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6144a7dc3881SStefano Zampini /* now compute set in local ordering */ 6145a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6146a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6147a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6148a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6149a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6150ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6151e456f2a8SStefano Zampini lsize++; 6152e456f2a8SStefano Zampini } 6153e456f2a8SStefano Zampini } 6154854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6155a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6156ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6157e456f2a8SStefano Zampini idxs[lsize++] = i; 6158e456f2a8SStefano Zampini } 6159e456f2a8SStefano Zampini } 6160a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6161a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6162e456f2a8SStefano Zampini *localis = localis_t; 6163e456f2a8SStefano Zampini PetscFunctionReturn(0); 6164e456f2a8SStefano Zampini } 6165906d46d4SStefano Zampini 6166b96c3477SStefano Zampini #undef __FUNCT__ 6167b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 616808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6169b96c3477SStefano Zampini { 6170a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6171b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6172b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6173a64f4aa4SStefano Zampini Mat S_j; 6174b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6175b96c3477SStefano Zampini PetscBool free_used_adj; 6176b96c3477SStefano Zampini PetscErrorCode ierr; 6177b96c3477SStefano Zampini 6178b96c3477SStefano Zampini PetscFunctionBegin; 6179b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6180b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 618108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6182b96c3477SStefano Zampini used_xadj = NULL; 6183b96c3477SStefano Zampini used_adjncy = NULL; 6184b96c3477SStefano Zampini } else { 618508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 618608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 618708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 618808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6189b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6190b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6191b96c3477SStefano Zampini } else { 61922fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6193b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6194b96c3477SStefano Zampini PetscInt nvtxs; 6195b96c3477SStefano Zampini 61962fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 61972fffb893SStefano Zampini if (flg_row) { 6198b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6199b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6200b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6201b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62022fffb893SStefano Zampini } else { 62032fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62042fffb893SStefano Zampini used_xadj = NULL; 62052fffb893SStefano Zampini used_adjncy = NULL; 62062fffb893SStefano Zampini } 62072fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6208b96c3477SStefano Zampini } 6209b96c3477SStefano Zampini } 6210d5574798SStefano Zampini 6211d5574798SStefano Zampini /* setup sub_schurs data */ 6212a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6213df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6214df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6215a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 621672b8c272SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 6217a64f4aa4SStefano Zampini } else { 62186816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 621904708bb6SStefano Zampini PetscBool isseqaij; 6220a3df083aSStefano Zampini PetscInt benign_n; 622172b8c272SStefano Zampini Mat change = NULL; 622272b8c272SStefano Zampini IS change_primal = NULL; 6223a3df083aSStefano Zampini 62245feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62255feab87aSStefano Zampini PetscInt n_vertices; 62265feab87aSStefano Zampini 62275feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62282034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62295feab87aSStefano Zampini } 623004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 623104708bb6SStefano Zampini if (!isseqaij) { 623204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 623304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 623404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 623504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 623604708bb6SStefano Zampini } else { 6237511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 623804708bb6SStefano Zampini } 623904708bb6SStefano Zampini } 6240a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6241a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6242ca92afb2SStefano Zampini } else { 6243a3df083aSStefano Zampini benign_n = 0; 6244ca92afb2SStefano Zampini } 6245bf3a8328SStefano Zampini /* If the user defines additional constraints, we import them here. 6246bf3a8328SStefano 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 */ 624772b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 624888c03ad3SStefano Zampini PC_IS *pcisf; 624988c03ad3SStefano Zampini PC_BDDC *pcbddcf; 625088c03ad3SStefano Zampini PC pcf; 625188c03ad3SStefano Zampini 625288c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 625388c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 625488c03ad3SStefano Zampini } 625588c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 625688c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 625788c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 625888c03ad3SStefano Zampini /* hacks */ 625988c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 626072b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 626172b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 626272b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 626372b8c272SStefano Zampini pcisf->n = pcis->n; 626472b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 626588c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 626688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 626788c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 626888c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 626988c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 627088c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 627172b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 627288c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 627388c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 627472b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 627572b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 627672b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 627772b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 627872b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 627988c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 628072b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 628188c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 628288c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 628388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 628488c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 628588c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 628688c03ad3SStefano Zampini pcf->ops->destroy = NULL; 628788c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 628888c03ad3SStefano Zampini } 628972b8c272SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 629072b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 629172b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6292ca92afb2SStefano Zampini } 6293d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6294b96c3477SStefano Zampini 6295b96c3477SStefano Zampini /* free adjacency */ 6296b96c3477SStefano Zampini if (free_used_adj) { 6297b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6298b96c3477SStefano Zampini } 6299b96c3477SStefano Zampini PetscFunctionReturn(0); 6300b96c3477SStefano Zampini } 6301b96c3477SStefano Zampini 6302b96c3477SStefano Zampini #undef __FUNCT__ 6303b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 630408122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6305b96c3477SStefano Zampini { 6306b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6307b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6308b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6309b96c3477SStefano Zampini PCBDDCGraph graph; 6310b96c3477SStefano Zampini PetscErrorCode ierr; 6311b96c3477SStefano Zampini 6312b96c3477SStefano Zampini PetscFunctionBegin; 6313b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 631408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63153301b35fSStefano Zampini IS verticesIS,verticescomm; 63163301b35fSStefano Zampini PetscInt vsize,*idxs; 6317b96c3477SStefano Zampini 6318b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63193301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63203301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63213301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63223301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63233301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6324b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63257fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6326441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63273301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6328b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6329b96c3477SStefano Zampini /* 6330b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6331b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6332b96c3477SStefano Zampini } 6333b96c3477SStefano Zampini */ 6334b96c3477SStefano Zampini } else { 6335b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6336b96c3477SStefano Zampini } 6337b96c3477SStefano Zampini 6338b96c3477SStefano Zampini /* sub_schurs init */ 6339a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6340a64f4aa4SStefano Zampini 6341b96c3477SStefano Zampini /* free graph struct */ 634208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6343b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6344b96c3477SStefano Zampini } 6345b96c3477SStefano Zampini PetscFunctionReturn(0); 6346b96c3477SStefano Zampini } 6347fa34dd3eSStefano Zampini 6348fa34dd3eSStefano Zampini #undef __FUNCT__ 6349fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6350fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6351fa34dd3eSStefano Zampini { 6352fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6353fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6354fa34dd3eSStefano Zampini PetscErrorCode ierr; 6355fa34dd3eSStefano Zampini 6356fa34dd3eSStefano Zampini PetscFunctionBegin; 6357fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6358fa34dd3eSStefano Zampini IS zerodiag = NULL; 63594f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6360fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 63614f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 636275c01103SStefano Zampini PetscReal norm; 6363fa34dd3eSStefano Zampini PetscInt i; 6364fa34dd3eSStefano Zampini 6365fa34dd3eSStefano Zampini /* B0 and B0_B */ 6366fa34dd3eSStefano Zampini if (zerodiag) { 6367fa34dd3eSStefano Zampini IS dummy; 6368fa34dd3eSStefano Zampini 63694f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 63704f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6371fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6372fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6373fa34dd3eSStefano Zampini } 6374fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6375fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6376fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6377fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6378fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6379fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6380fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6381fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6382fa34dd3eSStefano Zampini /* S_j */ 6383fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6384fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6385fa34dd3eSStefano Zampini 6386fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6387fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6388fa34dd3eSStefano Zampini /* continuous in primal space */ 6389fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6390fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6391fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6392fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 63934f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 63944f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6395fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6396fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6397fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6398fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6399fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6400fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6401fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6402fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6403fa34dd3eSStefano Zampini 6404fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6405fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6406fa34dd3eSStefano Zampini /* local with Schur */ 6407fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6408fa34dd3eSStefano Zampini if (zerodiag) { 6409fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64104f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6411fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6412fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6413fa34dd3eSStefano Zampini } 6414fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6415fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6416fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6417fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6418fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6419fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6420fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6421fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6422fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6423fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6424fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6425fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6426fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6427fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6428fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6429fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6430fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6431fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6432fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6433fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6434fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6435fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6436fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6437fa34dd3eSStefano Zampini if (zerodiag) { 6438fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6439fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6441fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6442fa34dd3eSStefano Zampini } 6443fa34dd3eSStefano Zampini /* BDDC */ 6444fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6445fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6446fa34dd3eSStefano Zampini 6447fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6448fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6449fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6450fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64514f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64524f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6453fa34dd3eSStefano Zampini } 64544f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6455fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6456fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6457fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6458fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6459fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6460fa34dd3eSStefano Zampini } 6461fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6462fa34dd3eSStefano Zampini } 6463