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); 315*9d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_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; 942af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 943af25d912SStefano 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; 1064*9d54b7f4SStefano Zampini PetscReal upper,lower; 1065862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 106608122e43SStefano Zampini PetscBLASInt B_N; 1067aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 106808122e43SStefano Zampini 1069*9d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1070*9d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 1071*9d54b7f4SStefano Zampini lower = thresh; 1072*9d54b7f4SStefano Zampini } else { 1073*9d54b7f4SStefano Zampini upper = 1./thresh; 1074*9d54b7f4SStefano Zampini lower = 0.; 1075*9d54b7f4SStefano Zampini } 1076862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1077ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1078f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1079f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1081aff50787SStefano Zampini PetscInt j,k; 1082aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1083aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1084aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 108508122e43SStefano Zampini } 108608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1087aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1088aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1089aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1090aff50787SStefano Zampini } 109108122e43SStefano Zampini } 109208122e43SStefano Zampini } else { 109308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109508122e43SStefano Zampini } 10968bec7fa6SStefano Zampini } else { 1097f6f667cfSStefano Zampini S = Sarray + cumarray; 1098f6f667cfSStefano Zampini St = Starray + cumarray; 10998bec7fa6SStefano Zampini } 1100aff50787SStefano Zampini /* see if we can save some work */ 1101aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1102aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1103aff50787SStefano Zampini } 1104aff50787SStefano Zampini 1105aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1106aff50787SStefano Zampini B_neigs = 0; 1107aff50787SStefano Zampini } else { 11089ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 110908122e43SStefano Zampini PetscBLASInt B_itype = 1; 1110f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11114c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11129552c7c7SStefano Zampini PetscInt nmin_s; 111308122e43SStefano Zampini 1114fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11158bec7fa6SStefano 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]]); 1116fd14bc51SStefano Zampini } 1117d16cbb6bSStefano Zampini 111808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1119d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1120d16cbb6bSStefano Zampini 1121d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 112208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1123*9d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 112408122e43SStefano Zampini #else 1125*9d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 112608122e43SStefano Zampini #endif 1127d16cbb6bSStefano Zampini } else { 1128d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1129d16cbb6bSStefano Zampini B_IL = 1; 1130d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1131*9d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1132d16cbb6bSStefano Zampini #else 1133*9d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1134d16cbb6bSStefano Zampini #endif 1135d16cbb6bSStefano Zampini } 113608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 113708122e43SStefano Zampini if (B_ierr) { 11386c4ed002SBarry 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); 11396c4ed002SBarry 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); 11406c4ed002SBarry 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); 114108122e43SStefano Zampini } 114208122e43SStefano Zampini 114308122e43SStefano Zampini if (B_neigs > nmax) { 1144fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1145fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1146fd14bc51SStefano Zampini } 1147*9d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 114808122e43SStefano Zampini B_neigs = nmax; 114908122e43SStefano Zampini } 115008122e43SStefano Zampini 11519552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11529552c7c7SStefano Zampini if (B_neigs < nmin_s) { 115308122e43SStefano Zampini PetscBLASInt B_neigs2; 115408122e43SStefano Zampini 1155*9d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1156f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1157*9d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 1158*9d54b7f4SStefano Zampini } else { 1159*9d54b7f4SStefano Zampini B_IL = B_neigs + 1; 1160*9d54b7f4SStefano Zampini B_IU = nmin_s; 1161*9d54b7f4SStefano Zampini } 1162fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1163fd14bc51SStefano 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); 1164fd14bc51SStefano Zampini } 11659ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 116608122e43SStefano Zampini PetscInt j; 116708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 116808122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 116908122e43SStefano Zampini } 117008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 117108122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 117208122e43SStefano Zampini } 117308122e43SStefano Zampini } else { 117408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 117508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 117608122e43SStefano Zampini } 117708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 117808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1179*9d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 118008122e43SStefano Zampini #else 1181*9d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 118208122e43SStefano Zampini #endif 118308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 118408122e43SStefano Zampini B_neigs += B_neigs2; 118508122e43SStefano Zampini } 118608122e43SStefano Zampini if (B_ierr) { 11876c4ed002SBarry 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); 11886c4ed002SBarry 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); 11896c4ed002SBarry 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); 119008122e43SStefano Zampini } 1191fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1192ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 119308122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 119408122e43SStefano Zampini if (eigs[j] == 0.0) { 1195ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 119608122e43SStefano Zampini } else { 1197*9d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1198ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1199*9d54b7f4SStefano Zampini } else { 1200*9d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 1201*9d54b7f4SStefano Zampini } 1202fd14bc51SStefano Zampini } 120308122e43SStefano Zampini } 120408122e43SStefano Zampini } 120508122e43SStefano Zampini } else { 120608122e43SStefano Zampini /* TODO */ 120708122e43SStefano Zampini } 1208aff50787SStefano Zampini } 12096c3e6151SStefano Zampini /* change the basis back to the original one */ 12106c3e6151SStefano Zampini if (sub_schurs->change) { 121172b8c272SStefano Zampini Mat change,phi,phit; 12126c3e6151SStefano Zampini 12136c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12146c3e6151SStefano Zampini PetscInt ii; 12156c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12166c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12176c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12186c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12196c3e6151SStefano Zampini } 12206c3e6151SStefano Zampini } 12216c3e6151SStefano Zampini } 122272b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12236c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 122472b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12256c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12266c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12276c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12286c3e6151SStefano Zampini } 12298bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12308bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12319162d606SStefano Zampini if (B_neigs) { 12329162d606SStefano 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); 1233fd14bc51SStefano Zampini 1234fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12359552c7c7SStefano Zampini PetscInt ii; 12369552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1237ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12389552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1239ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1240ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1241ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1242ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1243ac47001eSStefano Zampini #else 1244ac47001eSStefano 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); 1245ac47001eSStefano Zampini #endif 12469552c7c7SStefano Zampini } 12479552c7c7SStefano Zampini } 1248fd14bc51SStefano Zampini } 12499162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12509162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12519162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12529162d606SStefano Zampini cum++; 125308122e43SStefano Zampini } 125408122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 125508122e43SStefano Zampini /* shift for next computation */ 125608122e43SStefano Zampini cumarray += subset_size*subset_size; 125708122e43SStefano Zampini } 1258fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1259fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1260fd14bc51SStefano Zampini } 126108122e43SStefano Zampini 126208122e43SStefano Zampini if (mss) { 126308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 126408122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1265f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1266f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1267f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 126808122e43SStefano Zampini } 1269f6f667cfSStefano Zampini if (allocated_S_St) { 1270f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1271f6f667cfSStefano Zampini } 1272f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 127308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 127408122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 127508122e43SStefano Zampini #endif 127608122e43SStefano Zampini if (pcbddc->dbg_flag) { 12771b968477SStefano Zampini PetscInt maxneigs_r; 1278b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12799b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 128008122e43SStefano Zampini } 128108122e43SStefano Zampini PetscFunctionReturn(0); 128208122e43SStefano Zampini } 1283b1b3d7a2SStefano Zampini 1284674ae819SStefano Zampini #undef __FUNCT__ 1285c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1286c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1287c8587f34SStefano Zampini { 1288c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12898629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1290c8587f34SStefano Zampini PetscErrorCode ierr; 1291c8587f34SStefano Zampini 1292c8587f34SStefano Zampini PetscFunctionBegin; 1293f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12945e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1295c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1296c8587f34SStefano Zampini 1297684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12980fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1299684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1300c8587f34SStefano Zampini 1301c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1302b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1303c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1304c8587f34SStefano Zampini } 1305c8587f34SStefano Zampini 13068629588bSStefano Zampini /* 13078629588bSStefano Zampini Setup local correction and local part of coarse basis. 13088629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13098629588bSStefano Zampini */ 131047f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13118629588bSStefano Zampini 13128629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13138629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13148629588bSStefano Zampini 13158629588bSStefano Zampini /* free */ 13168629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1317c8587f34SStefano Zampini PetscFunctionReturn(0); 1318c8587f34SStefano Zampini } 1319c8587f34SStefano Zampini 1320c8587f34SStefano Zampini #undef __FUNCT__ 1321674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1322674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1323674ae819SStefano Zampini { 1324674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1325674ae819SStefano Zampini PetscErrorCode ierr; 1326674ae819SStefano Zampini 1327674ae819SStefano Zampini PetscFunctionBegin; 1328674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1329674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 133030368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1331674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1332674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1333785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1334674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1335f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1336f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1337785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 133863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 133963602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1340674ae819SStefano Zampini PetscFunctionReturn(0); 1341674ae819SStefano Zampini } 1342674ae819SStefano Zampini 1343674ae819SStefano Zampini #undef __FUNCT__ 1344674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1345674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1346674ae819SStefano Zampini { 1347674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13484f1b2e48SStefano Zampini PetscInt i; 1349674ae819SStefano Zampini PetscErrorCode ierr; 1350674ae819SStefano Zampini 1351674ae819SStefano Zampini PetscFunctionBegin; 1352b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1353674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 135416909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 13551dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1356674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1357674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13584f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13594f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13604f1b2e48SStefano Zampini } 13614f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1362b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1363674ae819SStefano Zampini PetscFunctionReturn(0); 1364674ae819SStefano Zampini } 1365674ae819SStefano Zampini 1366674ae819SStefano Zampini #undef __FUNCT__ 1367674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1368674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1369674ae819SStefano Zampini { 1370674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1371674ae819SStefano Zampini PetscErrorCode ierr; 1372674ae819SStefano Zampini 1373674ae819SStefano Zampini PetscFunctionBegin; 1374674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 137558da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1376ca92afb2SStefano Zampini PetscScalar *array; 137706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 137806656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 137958da7f69SStefano Zampini } 1380674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1381674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 138215aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 138315aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1384674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1385674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1386674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 138706656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1388674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1389674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13908ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1391674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1392674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1393674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1394f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1395f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1396f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1397f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1398727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13990e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1400f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 140170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 14026e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 140381d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14040369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14051dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 14064f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14078b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1408ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1409ca92afb2SStefano Zampini PetscInt i; 1410ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1411ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1412ca92afb2SStefano Zampini } 1413ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1414ca92afb2SStefano Zampini } 14154f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1416674ae819SStefano Zampini PetscFunctionReturn(0); 1417674ae819SStefano Zampini } 1418674ae819SStefano Zampini 1419674ae819SStefano Zampini #undef __FUNCT__ 1420f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1421f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14226bfb1811SStefano Zampini { 14236bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14246bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14256bfb1811SStefano Zampini VecType impVecType; 14264f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14276bfb1811SStefano Zampini PetscErrorCode ierr; 14286bfb1811SStefano Zampini 14296bfb1811SStefano Zampini PetscFunctionBegin; 14306c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1431e7b262bdSStefano Zampini /* get sizes */ 14324f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1433b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14346bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1435e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1436e7b262bdSStefano Zampini /* R nodes */ 1437e7b262bdSStefano Zampini old_size = -1; 1438e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1439e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1440e7b262bdSStefano Zampini } 1441e7b262bdSStefano Zampini if (n_R != old_size) { 1442e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1443e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14446bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14456bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14466bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14476bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1448e7b262bdSStefano Zampini } 1449e7b262bdSStefano Zampini /* local primal dofs */ 1450e7b262bdSStefano Zampini old_size = -1; 1451e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1452e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1453e7b262bdSStefano Zampini } 1454e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1455e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 145683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1457e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14586bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1459e7b262bdSStefano Zampini } 1460e7b262bdSStefano Zampini /* local explicit constraints */ 1461e7b262bdSStefano Zampini old_size = -1; 1462e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1463e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1464e7b262bdSStefano Zampini } 1465e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1466e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 146783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 146883b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 146983b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 147083b7ccabSStefano Zampini } 14716bfb1811SStefano Zampini PetscFunctionReturn(0); 14726bfb1811SStefano Zampini } 14736bfb1811SStefano Zampini 14746bfb1811SStefano Zampini #undef __FUNCT__ 147547f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 147647f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 147788ebb749SStefano Zampini { 147825084f0cSStefano Zampini PetscErrorCode ierr; 147925084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 148088ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 148188ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1482d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 148325084f0cSStefano Zampini /* submatrices of local problem */ 148480677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 148506656605SStefano Zampini /* submatrices of local coarse problem */ 148606656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 148725084f0cSStefano Zampini /* working matrices */ 148806656605SStefano Zampini Mat C_CR; 148925084f0cSStefano Zampini /* additional working stuff */ 149006656605SStefano Zampini PC pc_R; 14914f1b2e48SStefano Zampini Mat F; 14925cbda25cSStefano Zampini Vec dummy_vec; 1493a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 149425084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 149506656605SStefano Zampini PetscScalar *work; 149606656605SStefano Zampini PetscInt *idx_V_B; 1497ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 149806656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1499ffd830a3SStefano Zampini 150025084f0cSStefano Zampini /* some shortcuts to scalars */ 150106656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 150288ebb749SStefano Zampini 150388ebb749SStefano Zampini PetscFunctionBegin; 1504ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1505ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1506ffd830a3SStefano Zampini } 1507ffd830a3SStefano Zampini 1508ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1509b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15104f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1511b371cd4fSStefano Zampini n_B = pcis->n_B; 1512b371cd4fSStefano Zampini n_D = pcis->n - n_B; 151388ebb749SStefano Zampini n_R = pcis->n - n_vertices; 151488ebb749SStefano Zampini 151588ebb749SStefano Zampini /* vertices in boundary numbering */ 1516785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15170e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 15186c4ed002SBarry 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); 151988ebb749SStefano Zampini 152006656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1521019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 152206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 152306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 152406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 152506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 152606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 152706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 152806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 152906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 153006656605SStefano Zampini 153106656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 153206656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 153306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 153406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 153506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1536ffd830a3SStefano Zampini lda_rhs = n_R; 1537a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 153806656605SStefano Zampini if (isLU || isILU || isCHOL) { 153906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1540df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1541df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1542d62866d3SStefano Zampini MatFactorType type; 1543d62866d3SStefano Zampini 1544df4d28bfSStefano Zampini F = reuse_solver->F; 15456816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1546d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1547ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1548df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 154906656605SStefano Zampini } else { 155006656605SStefano Zampini F = NULL; 155106656605SStefano Zampini } 155206656605SStefano Zampini 1553ffd830a3SStefano Zampini /* allocate workspace */ 1554ffd830a3SStefano Zampini n = 0; 1555ffd830a3SStefano Zampini if (n_constraints) { 1556ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1557ffd830a3SStefano Zampini } 1558ffd830a3SStefano Zampini if (n_vertices) { 1559ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1560ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1561ffd830a3SStefano Zampini } 1562ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1563ffd830a3SStefano Zampini 15645cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 15655cbda25cSStefano Zampini dummy_vec = NULL; 15665cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 15675cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 15685cbda25cSStefano Zampini } 15695cbda25cSStefano Zampini 157088ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 157188ebb749SStefano Zampini if (n_constraints) { 157272b8c272SStefano Zampini Mat M1,M2,M3,C_B; 157306656605SStefano Zampini IS is_aux; 157480677318SStefano Zampini PetscScalar *array,*array2; 157506656605SStefano Zampini 1576f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 157780677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 157888ebb749SStefano Zampini 157925084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 158025084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15818ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 158272b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 158388ebb749SStefano Zampini 158480677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 158580677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1586ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 158788ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 158806656605SStefano Zampini const PetscScalar *row_cmat_values; 158906656605SStefano Zampini const PetscInt *row_cmat_indices; 159006656605SStefano Zampini PetscInt size_of_constraint,j; 159188ebb749SStefano Zampini 159206656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 159306656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1594ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 159506656605SStefano Zampini } 159606656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 159706656605SStefano Zampini } 1598ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 159906656605SStefano Zampini if (F) { 160006656605SStefano Zampini Mat B; 160106656605SStefano Zampini 1602ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1603a3df083aSStefano Zampini if (need_benign_correction) { 1604df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1605a3df083aSStefano Zampini 160672b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 160772b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1608a3df083aSStefano Zampini } 160980677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1610a3df083aSStefano Zampini if (need_benign_correction) { 1611a3df083aSStefano Zampini PetscScalar *marr; 1612df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1613a3df083aSStefano Zampini 1614a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 16155cbda25cSStefano Zampini if (lda_rhs != n_R) { 16165cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 16175cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 16185cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 16195cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 16205cbda25cSStefano Zampini } 16215cbda25cSStefano Zampini } else { 1622a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1623a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 16245cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1625a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1626a3df083aSStefano Zampini } 16275cbda25cSStefano Zampini } 1628a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1629a3df083aSStefano Zampini } 163006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 163106656605SStefano Zampini } else { 163280677318SStefano Zampini PetscScalar *marr; 163380677318SStefano Zampini 163480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 163506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1636ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1637ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 163806656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 163906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 164006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 164106656605SStefano Zampini } 164280677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 164306656605SStefano Zampini } 164480677318SStefano Zampini if (!pcbddc->switch_static) { 164580677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 164680677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 164780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 164880677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1649ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 165080677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 165180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 165480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 165580677318SStefano Zampini } 165680677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 165780677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 165872b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 165980677318SStefano Zampini } else { 1660ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1661ffd830a3SStefano Zampini IS dummy; 1662ffd830a3SStefano Zampini 1663ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 166472b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1665ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1666ffd830a3SStefano Zampini } else { 166780677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 166880677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1669ffd830a3SStefano Zampini } 167025084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 167180677318SStefano Zampini } 167280677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 167380677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 167480677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 167506656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 167606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 167780677318SStefano Zampini if (isCHOL) { 167880677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 167980677318SStefano Zampini } else { 168025084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 168180677318SStefano Zampini } 168280677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 168306656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 168425084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 168525084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 168625084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 168780677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 168872b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 168972b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 169006656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 169106656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1692f4ddd8eeSStefano Zampini } 1693fc227af8SStefano Zampini 1694fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 169588ebb749SStefano Zampini if (n_vertices) { 169606656605SStefano Zampini IS is_aux; 16973a50541eSStefano Zampini 1698df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16996816873aSStefano Zampini IS tis; 17006816873aSStefano Zampini 17016816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17026816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17036816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17046816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17056816873aSStefano Zampini } else { 17063a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17076816873aSStefano Zampini } 17089577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17099577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 171004708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 171125084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 171288ebb749SStefano Zampini } 171388ebb749SStefano Zampini 171488ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1715f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 171606656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 171706656605SStefano Zampini if (pcbddc->coarse_phi_D) { 171806656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 171906656605SStefano Zampini } 1720f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 172106656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 172206656605SStefano Zampini PetscScalar *marray; 172306656605SStefano Zampini 172406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 172506656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1726f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1727f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1728f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1729f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1730f4ddd8eeSStefano Zampini } 1731f4ddd8eeSStefano Zampini } 173206656605SStefano Zampini 1733f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 173406656605SStefano Zampini PetscScalar *marray; 173588ebb749SStefano Zampini 173606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17378eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 173806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 173988ebb749SStefano Zampini } 17403301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 174106656605SStefano Zampini n *= 2; 174288ebb749SStefano Zampini } 174306656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 174406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 174506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17468eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 174706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 174806656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 174988ebb749SStefano Zampini } 17503301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 175106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17528eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 175306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 175406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 175588ebb749SStefano Zampini } 175688ebb749SStefano Zampini } else { 1757c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1758c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17591b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1760c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1761c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1762c0553b1fSStefano Zampini } 176388ebb749SStefano Zampini } 176406656605SStefano Zampini } 1765019a44ceSStefano Zampini 176606656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17674f1b2e48SStefano Zampini p0_lidx_I = NULL; 17684f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1769d12edf2fSStefano Zampini const PetscInt *idxs; 1770d12edf2fSStefano Zampini 1771d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17724f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17734f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17744f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17754f1b2e48SStefano Zampini } 1776d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1777d12edf2fSStefano Zampini } 1778d16cbb6bSStefano Zampini 177906656605SStefano Zampini /* vertices */ 178006656605SStefano Zampini if (n_vertices) { 178116f15bc4SStefano Zampini 1782af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 178304708bb6SStefano Zampini 178416f15bc4SStefano Zampini if (n_R) { 178514393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 178606656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 178716f15bc4SStefano Zampini PetscScalar *x,*y; 178804708bb6SStefano Zampini PetscBool isseqaij; 178906656605SStefano Zampini 179021eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 179114393ed6SStefano Zampini if (need_benign_correction) { 179214393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 179314393ed6SStefano Zampini IS is_p0; 179414393ed6SStefano Zampini PetscInt *idxs_p0,n; 179514393ed6SStefano Zampini 179614393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 179714393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 179814393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 1799af25d912SStefano 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); 180014393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 180114393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 180214393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 180314393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 180414393ed6SStefano Zampini } 180514393ed6SStefano Zampini 1806ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1807af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1808ffd830a3SStefano Zampini } else { 1809ca92afb2SStefano Zampini PetscScalar *av,*array; 1810ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1811ca92afb2SStefano Zampini PetscInt n; 1812ca92afb2SStefano Zampini PetscBool flg_row; 1813ffd830a3SStefano Zampini 1814ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1815ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1816*9d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1817ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1818ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1819ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1820ca92afb2SStefano Zampini PetscInt j; 1821ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1822ffd830a3SStefano Zampini } 1823ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1824ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1825ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1826ffd830a3SStefano Zampini } 1827ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1828a3df083aSStefano Zampini if (need_benign_correction) { 1829df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1830a3df083aSStefano Zampini PetscScalar *marr; 1831a3df083aSStefano Zampini 1832a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 183314393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 183414393ed6SStefano Zampini 183514393ed6SStefano Zampini | 0 0 0 | (V) 183614393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 183714393ed6SStefano Zampini | 0 0 -1 | (p0) 183814393ed6SStefano Zampini 183914393ed6SStefano Zampini */ 1840df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 184114393ed6SStefano Zampini const PetscScalar *vals; 184214393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 184314393ed6SStefano Zampini PetscInt n,j,nz; 184414393ed6SStefano Zampini 1845df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1846df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 184714393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 184814393ed6SStefano Zampini for (j=0;j<n;j++) { 184914393ed6SStefano Zampini PetscScalar val = vals[j]; 185014393ed6SStefano Zampini PetscInt k,col = idxs[j]; 185114393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 185214393ed6SStefano Zampini } 185314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1854df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 185514393ed6SStefano Zampini } 185672b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 185772b8c272SStefano Zampini } 185872b8c272SStefano Zampini if (F) { 185914393ed6SStefano Zampini /* need to correct the rhs */ 186072b8c272SStefano Zampini if (need_benign_correction) { 186172b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 186272b8c272SStefano Zampini PetscScalar *marr; 186372b8c272SStefano Zampini 186472b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 18655cbda25cSStefano Zampini if (lda_rhs != n_R) { 18665cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18675cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18685cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 18695cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18705cbda25cSStefano Zampini } 18715cbda25cSStefano Zampini } else { 1872a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1873a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18745cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1875a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1876a3df083aSStefano Zampini } 18775cbda25cSStefano Zampini } 1878a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1879a3df083aSStefano Zampini } 188006656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 188114393ed6SStefano Zampini /* need to correct the solution */ 1882a3df083aSStefano Zampini if (need_benign_correction) { 1883df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1884a3df083aSStefano Zampini PetscScalar *marr; 1885a3df083aSStefano Zampini 1886a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 18875cbda25cSStefano Zampini if (lda_rhs != n_R) { 18885cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18895cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18905cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 18915cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18925cbda25cSStefano Zampini } 18935cbda25cSStefano Zampini } else { 1894a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1895a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18965cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1897a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1898a3df083aSStefano Zampini } 18995cbda25cSStefano Zampini } 1900a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1901a3df083aSStefano Zampini } 190206656605SStefano Zampini } else { 190306656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 190406656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1905ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1906ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 190706656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 190806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 190906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 191006656605SStefano Zampini } 191106656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 191206656605SStefano Zampini } 191380677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1914ffd830a3SStefano Zampini /* S_VV and S_CV */ 191506656605SStefano Zampini if (n_constraints) { 191606656605SStefano Zampini Mat B; 191780677318SStefano Zampini 1918ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 191980677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1920ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1921ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 192280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 192580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 192680677318SStefano Zampini } 1927ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 192880677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 192980677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1930ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 193180677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 193206656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1933ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1934ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 193506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 193606656605SStefano Zampini } 193704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 193804708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 1939511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 194004708bb6SStefano Zampini } 1941ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1942ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1943ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1944ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1945ffd830a3SStefano Zampini } 194606656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 194714393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 194814393ed6SStefano Zampini if (need_benign_correction) { 1949df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 195014393ed6SStefano Zampini PetscScalar *marr,*sums; 195114393ed6SStefano Zampini 195214393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 195314393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1954df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 195514393ed6SStefano Zampini const PetscScalar *vals; 195614393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 195714393ed6SStefano Zampini PetscInt n,j,nz; 195814393ed6SStefano Zampini 1959df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1960df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 196114393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 196214393ed6SStefano Zampini PetscInt k; 196314393ed6SStefano Zampini sums[j] = 0.; 196414393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 196514393ed6SStefano Zampini } 196614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 196714393ed6SStefano Zampini for (j=0;j<n;j++) { 196814393ed6SStefano Zampini PetscScalar val = vals[j]; 196914393ed6SStefano Zampini PetscInt k; 197014393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 197114393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 197214393ed6SStefano Zampini } 197314393ed6SStefano Zampini } 197414393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1975df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 197614393ed6SStefano Zampini } 197714393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 197814393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 197914393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 198014393ed6SStefano Zampini } 198180677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 198206656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 198306656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 198406656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 198506656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 198606656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 198706656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 198806656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1989d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1990019a44ceSStefano Zampini } else { 1991d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1992d16cbb6bSStefano Zampini } 199321eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1994d16cbb6bSStefano Zampini 199506656605SStefano Zampini /* coarse basis functions */ 199606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 199716f15bc4SStefano Zampini PetscScalar *y; 199816f15bc4SStefano Zampini 1999ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 200006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 200206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 200506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 200606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 200706656605SStefano Zampini 200806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20094f1b2e48SStefano Zampini PetscInt j; 20104f1b2e48SStefano Zampini 201106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 201206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 201306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20164f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 201706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 201806656605SStefano Zampini } 201906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 202006656605SStefano Zampini } 202104708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 202204708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 202306656605SStefano Zampini } 20245cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 202506656605SStefano Zampini 202606656605SStefano Zampini if (n_constraints) { 202706656605SStefano Zampini Mat B; 202806656605SStefano Zampini 2029ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 203006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 203180677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 203206656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 203306656605SStefano Zampini if (n_vertices) { 203480677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 203580677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 203680677318SStefano Zampini } else { 203780677318SStefano Zampini Mat S_VCt; 203880677318SStefano Zampini 2039ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2040ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 204172b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2042ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2043ffd830a3SStefano Zampini } 204480677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 204580677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 204680677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 204780677318SStefano Zampini } 204806656605SStefano Zampini } 204906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 205006656605SStefano Zampini /* coarse basis functions */ 205106656605SStefano Zampini for (i=0;i<n_constraints;i++) { 205206656605SStefano Zampini PetscScalar *y; 205306656605SStefano Zampini 2054ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 205506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 205606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 205706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 206006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 206106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20624f1b2e48SStefano Zampini PetscInt j; 20634f1b2e48SStefano Zampini 206406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 206506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 206606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20694f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 207006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 207106656605SStefano Zampini } 207206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 207306656605SStefano Zampini } 207406656605SStefano Zampini } 207580677318SStefano Zampini if (n_constraints) { 207680677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 207780677318SStefano Zampini } 20784f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 207972b8c272SStefano Zampini 208072b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 208172b8c272SStefano Zampini if (pcbddc->benign_n) { 208272b8c272SStefano Zampini Mat B0_B,B0_BPHI; 208372b8c272SStefano Zampini IS is_dummy; 208472b8c272SStefano Zampini PetscScalar *data; 208572b8c272SStefano Zampini PetscInt j; 208672b8c272SStefano Zampini 208772b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 208872b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 208972b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 209072b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 209186c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 209272b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 209372b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 209472b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 209572b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 209672b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 209772b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 209872b8c272SStefano Zampini } 209972b8c272SStefano Zampini } 210072b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 210172b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 210272b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 210372b8c272SStefano Zampini } 2104019a44ceSStefano Zampini 210506656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 21063301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2107ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2108ffd830a3SStefano Zampini PetscScalar *marray; 210906656605SStefano Zampini 211006656605SStefano Zampini if (n_constraints) { 2111ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 211206656605SStefano Zampini 2113af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 211406656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2115ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 211616f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 211706656605SStefano Zampini if (n_vertices) { 2118ffd830a3SStefano Zampini Mat S_VCT; 211906656605SStefano Zampini 212006656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2121ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 212216f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 212306656605SStefano Zampini } 2124ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 212506656605SStefano Zampini } 212616f15bc4SStefano Zampini if (n_vertices && n_R) { 2127ffd830a3SStefano Zampini PetscScalar *av,*marray; 2128ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2129ffd830a3SStefano Zampini PetscInt n; 2130ffd830a3SStefano Zampini PetscBool flg_row; 213106656605SStefano Zampini 2132ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2133af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2134ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2135ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2136ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2137ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2138ffd830a3SStefano Zampini PetscInt j; 2139ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2140ffd830a3SStefano Zampini } 2141ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2142ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2143ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 214406656605SStefano Zampini } 214506656605SStefano Zampini 2146ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2147ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2148ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2149ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2150ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 215106656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 215206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 215306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 215406656605SStefano Zampini } 2155ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2156ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2157ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2158ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2159ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2160ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2161ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2162ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 216306656605SStefano Zampini } 2164ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 216506656605SStefano Zampini /* coarse basis functions */ 216606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 216706656605SStefano Zampini PetscScalar *y; 216806656605SStefano Zampini 2169ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 217006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 217106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 217206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217406656605SStefano Zampini if (i<n_vertices) { 217506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 217606656605SStefano Zampini } 217706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 217806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 217906656605SStefano Zampini 218006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 218106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 218206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 218306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 218406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 218506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 218606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 218706656605SStefano Zampini } 218806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 218906656605SStefano Zampini } 2190ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2191ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 219206656605SStefano Zampini } 2193d62866d3SStefano Zampini /* free memory */ 219488ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 219506656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 219606656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 219706656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 219806656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2199d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2200d62866d3SStefano Zampini if (n_vertices) { 2201d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2202d62866d3SStefano Zampini } 2203d62866d3SStefano Zampini if (n_constraints) { 2204d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2205d62866d3SStefano Zampini } 220688ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 220788ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 220888ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2209d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 221088ebb749SStefano Zampini Mat coarse_sub_mat; 221125084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 221288ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 221388ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 221488ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22158bec7fa6SStefano Zampini Mat C_B,CPHI; 22168bec7fa6SStefano Zampini IS is_dummy; 22178bec7fa6SStefano Zampini Vec mones; 221888ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 221988ebb749SStefano Zampini PetscReal real_value; 222088ebb749SStefano Zampini 2221a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2222a3df083aSStefano Zampini Mat A; 2223a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2224a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2225a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2226a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2227a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2228a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2229a3df083aSStefano Zampini } else { 223088ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 223188ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 223288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 223388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2234a3df083aSStefano Zampini } 223588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 223688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2237ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 223888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 223988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 224088ebb749SStefano Zampini } 224188ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 224288ebb749SStefano Zampini 224325084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22443301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 224525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2246ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 224788ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 224888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 224988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 225088ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 225188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 225288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 225388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 225488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 225588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 225688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 225788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 225888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 225988ebb749SStefano Zampini } else { 226088ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 226188ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 226288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 226388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 226488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 226588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 226688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 226788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 226888ebb749SStefano Zampini } 226988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 227088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 227188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2272511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 22734f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2274fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2275d12edf2fSStefano Zampini PetscScalar *data,*data2; 22764f1b2e48SStefano Zampini PetscInt j; 2277d12edf2fSStefano Zampini 22784f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2279fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2280d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 228186c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2282d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2283d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 22844f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 22854f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2286d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 22874f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 22884f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22894f1b2e48SStefano Zampini } 2290d12edf2fSStefano Zampini } 2291d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2292d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2293d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2294d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2295d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2296d12edf2fSStefano Zampini } 2297d12edf2fSStefano Zampini #if 0 2298d12edf2fSStefano Zampini { 2299d12edf2fSStefano Zampini PetscViewer viewer; 2300d12edf2fSStefano Zampini char filename[256]; 2301ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2302d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2303d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2304ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2305ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2306ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2307d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 230872b8c272SStefano Zampini if (save_change) { 230972b8c272SStefano Zampini Mat phi_B; 231072b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 231172b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 231272b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 231372b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 231472b8c272SStefano Zampini } else { 2315ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2316ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 231772b8c272SStefano Zampini } 2318ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2319ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2320ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2321ffd830a3SStefano Zampini } 2322ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2323ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2324ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2325ffd830a3SStefano Zampini } 232672b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2327ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2328ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2329ffd830a3SStefano Zampini } 2330d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2331d12edf2fSStefano Zampini } 2332d12edf2fSStefano Zampini #endif 233381d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23348bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23351575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 233606656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23378bec7fa6SStefano Zampini 23388bec7fa6SStefano Zampini /* check constraints */ 2339a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2340a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 23414f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23428bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2343a00504b5SStefano Zampini } else { 2344a00504b5SStefano Zampini PetscScalar *data; 2345a00504b5SStefano Zampini Mat tmat; 2346a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2347a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2348a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2349a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2350a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2351a00504b5SStefano Zampini } 23528bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23538bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23548bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23558bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2356bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2357ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2358bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2359bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2360bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2361bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2362bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 236388ebb749SStefano Zampini } 23648bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23658bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23668bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23678bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 236825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 236988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 237088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 237188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 237288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 237388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 237488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 237588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 237688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 237788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 237888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2379ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 238088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 238188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 238288ebb749SStefano Zampini } 238388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 238488ebb749SStefano Zampini } 23858629588bSStefano Zampini /* get back data */ 23868629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 238788ebb749SStefano Zampini PetscFunctionReturn(0); 238888ebb749SStefano Zampini } 238988ebb749SStefano Zampini 239088ebb749SStefano Zampini #undef __FUNCT__ 2391d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2392d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2393aa0d41d4SStefano Zampini { 2394d65f70fdSStefano Zampini Mat *work_mat; 2395d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2396d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2397d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2398aa0d41d4SStefano Zampini PetscErrorCode ierr; 2399aa0d41d4SStefano Zampini 2400aa0d41d4SStefano Zampini PetscFunctionBegin; 2401d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2402d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2403d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2404d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2405aa0d41d4SStefano Zampini 2406d65f70fdSStefano Zampini if (!rsorted) { 2407906d46d4SStefano Zampini const PetscInt *idxs; 2408906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2409aa0d41d4SStefano Zampini 2410d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2411d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2412d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2413d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2414aa0d41d4SStefano Zampini } 2415d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2416d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2417d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2418d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2419aa0d41d4SStefano Zampini } 2420d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2421d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2422d65f70fdSStefano Zampini } else { 2423d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2424d65f70fdSStefano Zampini isrow_s = isrow; 2425aa0d41d4SStefano Zampini } 2426906d46d4SStefano Zampini 2427d65f70fdSStefano Zampini if (!csorted) { 2428d65f70fdSStefano Zampini if (isrow == iscol) { 2429d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2430d65f70fdSStefano Zampini iscol_s = isrow_s; 2431d65f70fdSStefano Zampini } else { 2432d65f70fdSStefano Zampini const PetscInt *idxs; 2433d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2434906d46d4SStefano Zampini 2435d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2436d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2437d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2438d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2439d65f70fdSStefano Zampini } 2440d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2441d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2442d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2443d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2444d65f70fdSStefano Zampini } 2445d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2446d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2447d65f70fdSStefano Zampini } 2448d65f70fdSStefano Zampini } else { 2449d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2450d65f70fdSStefano Zampini iscol_s = iscol; 2451d65f70fdSStefano Zampini } 2452d65f70fdSStefano Zampini 2453d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2454d65f70fdSStefano Zampini 2455d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2456906d46d4SStefano Zampini Mat new_mat; 2457d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2458906d46d4SStefano Zampini 2459d65f70fdSStefano Zampini if (!rsorted) { 2460d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2461d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2462d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2463d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2464906d46d4SStefano Zampini } 2465d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2466d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2467d65f70fdSStefano Zampini } else { 2468d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2469906d46d4SStefano Zampini } 2470d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2471d65f70fdSStefano Zampini 2472d65f70fdSStefano Zampini if (!csorted) { 2473d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2474d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2475d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2476d65f70fdSStefano Zampini } else { 2477d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2478d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2479d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2480d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2481d65f70fdSStefano Zampini } 2482d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2483d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2484d65f70fdSStefano Zampini } 2485d65f70fdSStefano Zampini } else { 2486d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2487d65f70fdSStefano Zampini } 2488d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2489d65f70fdSStefano Zampini 2490d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2491d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2492d65f70fdSStefano Zampini work_mat[0] = new_mat; 2493d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2494d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2495d65f70fdSStefano Zampini } 2496d65f70fdSStefano Zampini 2497d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2498d65f70fdSStefano Zampini *B = work_mat[0]; 2499d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2500d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2501d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2502d65f70fdSStefano Zampini PetscFunctionReturn(0); 2503d65f70fdSStefano Zampini } 2504d65f70fdSStefano Zampini 2505d65f70fdSStefano Zampini #undef __FUNCT__ 25065e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 25075e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2508aa0d41d4SStefano Zampini { 2509aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 25105e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2511d65f70fdSStefano Zampini Mat new_mat; 25125e8657edSStefano Zampini IS is_local,is_global; 2513d65f70fdSStefano Zampini PetscInt local_size; 2514d65f70fdSStefano Zampini PetscBool isseqaij; 2515aa0d41d4SStefano Zampini PetscErrorCode ierr; 2516aa0d41d4SStefano Zampini 2517aa0d41d4SStefano Zampini PetscFunctionBegin; 2518aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25195e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25205e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2521b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2522aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2523d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2524aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2525906d46d4SStefano Zampini 2526906d46d4SStefano Zampini /* check */ 2527906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2528906d46d4SStefano Zampini Vec x,x_change; 2529906d46d4SStefano Zampini PetscReal error; 2530906d46d4SStefano Zampini 25315e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2532906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25335e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2534e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2535e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2536d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2537e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2538e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2539906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2540906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2541906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2542906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2543906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2544906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2545906d46d4SStefano Zampini } 2546906d46d4SStefano Zampini 254722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25489b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 254922d5777bSStefano Zampini if (isseqaij) { 2550a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2551a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2552aa0d41d4SStefano Zampini } else { 2553a00504b5SStefano Zampini Mat work_mat; 25541cf9b237SStefano Zampini 2555a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2556aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2557a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2558aa0d41d4SStefano Zampini } 25593301b35fSStefano Zampini if (matis->A->symmetric_set) { 25603301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2561e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25623301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2563e496cd5dSStefano Zampini #endif 25643301b35fSStefano Zampini } 2565d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2566aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2567aa0d41d4SStefano Zampini } 2568aa0d41d4SStefano Zampini 2569aa0d41d4SStefano Zampini #undef __FUNCT__ 2570a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25718ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2572a64d13efSStefano Zampini { 2573a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2574a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2575d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 257653892102SStefano Zampini PetscInt *idx_R_local=NULL; 25773a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25783a50541eSStefano Zampini PetscInt vbs,bs; 25796816873aSStefano Zampini PetscBT bitmask=NULL; 2580a64d13efSStefano Zampini PetscErrorCode ierr; 2581a64d13efSStefano Zampini 2582a64d13efSStefano Zampini PetscFunctionBegin; 2583b23d619eSStefano Zampini /* 2584b23d619eSStefano Zampini No need to setup local scatters if 2585b23d619eSStefano Zampini - primal space is unchanged 2586b23d619eSStefano Zampini AND 2587b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2588b23d619eSStefano Zampini AND 2589b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2590b23d619eSStefano Zampini */ 2591b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2592f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2593f4ddd8eeSStefano Zampini } 2594f4ddd8eeSStefano Zampini /* destroy old objects */ 2595f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2596f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2597f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2598a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2599b371cd4fSStefano Zampini n_B = pcis->n_B; 2600b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2601b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 26023a50541eSStefano Zampini 2603a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 26046816873aSStefano Zampini 260553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2606df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2607854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2608a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2609a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 26100e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2611a64d13efSStefano Zampini } 2612a64d13efSStefano Zampini 2613a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26144641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26156816873aSStefano Zampini idx_R_local[n_R++] = i; 2616a64d13efSStefano Zampini } 2617a64d13efSStefano Zampini } 2618df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2619df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26206816873aSStefano Zampini 2621df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2622df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26236816873aSStefano Zampini } 26243a50541eSStefano Zampini 26253a50541eSStefano Zampini /* Block code */ 26263a50541eSStefano Zampini vbs = 1; 26273a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26283a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26293a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26303a50541eSStefano Zampini PetscInt *vary; 2631df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2632785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26333a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2634d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2635d3df7717SStefano 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 */ 26360e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2637d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26383a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26393a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26403a50541eSStefano Zampini break; 26413a50541eSStefano Zampini } 26423a50541eSStefano Zampini } 2643d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2644d3df7717SStefano Zampini } else { 2645d3df7717SStefano Zampini /* Verify directly the R set */ 2646d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2647d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2648d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2649d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2650d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2651d3df7717SStefano Zampini break; 2652d3df7717SStefano Zampini } 2653d3df7717SStefano Zampini } 2654d3df7717SStefano Zampini } 2655d3df7717SStefano Zampini } 26563a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26573a50541eSStefano Zampini vbs = bs; 26583a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26593a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26603a50541eSStefano Zampini } 26613a50541eSStefano Zampini } 26623a50541eSStefano Zampini } 26633a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2664df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2665df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 266653892102SStefano Zampini 2667df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2668df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 266953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2670df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 267153892102SStefano Zampini } else { 26723a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 267353892102SStefano Zampini } 2674a64d13efSStefano Zampini 2675a64d13efSStefano Zampini /* print some info if requested */ 2676a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2677a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2678a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26791575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2680a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2681a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 26824f1b2e48SStefano 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); 2683a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2684a64d13efSStefano Zampini } 2685a64d13efSStefano Zampini 2686a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2687df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 26886816873aSStefano Zampini IS is_aux1,is_aux2; 26896816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 26906816873aSStefano Zampini 26913a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2692854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2693854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2694a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26954641a718SStefano Zampini for (i=0; i<n_D; i++) { 26964641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26974641a718SStefano Zampini } 2698a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2699a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27004641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 27014641a718SStefano Zampini aux_array1[j++] = i; 2702a64d13efSStefano Zampini } 2703a64d13efSStefano Zampini } 2704a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2705a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2706a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 27074641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 27084641a718SStefano Zampini aux_array2[j++] = i; 2709a64d13efSStefano Zampini } 2710a64d13efSStefano Zampini } 2711a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2712a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2713a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2714a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2715a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2716a64d13efSStefano Zampini 27178eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2718785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2719a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27204641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27214641a718SStefano Zampini aux_array1[j++] = i; 2722a64d13efSStefano Zampini } 2723a64d13efSStefano Zampini } 2724a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2725a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2726a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2727a64d13efSStefano Zampini } 27284641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27293a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2730d62866d3SStefano Zampini } else { 2731df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27326816873aSStefano Zampini IS tis; 27336816873aSStefano Zampini PetscInt schur_size; 27346816873aSStefano Zampini 2735df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27366816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2737df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27386816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27396816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27406816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27416816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27426816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2743d62866d3SStefano Zampini } 2744d62866d3SStefano Zampini } 2745a64d13efSStefano Zampini PetscFunctionReturn(0); 2746a64d13efSStefano Zampini } 2747a64d13efSStefano Zampini 2748304d26faSStefano Zampini 2749304d26faSStefano Zampini #undef __FUNCT__ 2750304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2751684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2752304d26faSStefano Zampini { 2753304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2754304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2755304d26faSStefano Zampini PC pc_temp; 2756304d26faSStefano Zampini Mat A_RR; 2757f4ddd8eeSStefano Zampini MatReuse reuse; 2758304d26faSStefano Zampini PetscScalar m_one = -1.0; 2759304d26faSStefano Zampini PetscReal value; 276004708bb6SStefano Zampini PetscInt n_D,n_R; 27619577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2762304d26faSStefano Zampini PetscErrorCode ierr; 2763e604994aSStefano Zampini /* prefixes stuff */ 2764312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2765e604994aSStefano Zampini size_t len; 2766304d26faSStefano Zampini 2767304d26faSStefano Zampini PetscFunctionBegin; 2768304d26faSStefano Zampini 2769e604994aSStefano Zampini /* compute prefixes */ 2770e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2771e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2772e604994aSStefano Zampini if (!pcbddc->current_level) { 2773e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2774e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2775e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2776e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2777e604994aSStefano Zampini } else { 2778e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2779312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2780e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2781e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2782312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2783312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 278434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 278534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2786e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2787e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2788e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2789e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2790e604994aSStefano Zampini } 2791e604994aSStefano Zampini 2792304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2793684f6988SStefano Zampini if (dirichlet) { 2794d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2795450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2796450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2797450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2798450f8f5eSStefano Zampini } 2799450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2800a3df083aSStefano Zampini Mat A_IIn; 2801a3df083aSStefano Zampini 2802a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2803a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2804a3df083aSStefano Zampini pcis->A_II = A_IIn; 2805a3df083aSStefano Zampini } 2806450f8f5eSStefano Zampini } 28073301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28083301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2809964fefecSStefano Zampini } 2810ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2811964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2812304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2813304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2814304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2815304d26faSStefano Zampini /* default */ 2816304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2817e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28189577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2819304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28209577ea80SStefano Zampini if (issbaij) { 28219577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28229577ea80SStefano Zampini } else { 2823304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28249577ea80SStefano Zampini } 2825304d26faSStefano Zampini /* Allow user's customization */ 2826304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2827304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2828304d26faSStefano Zampini } 2829d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2830df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2831df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2832d62866d3SStefano Zampini 2833df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2834d5574798SStefano Zampini } 2835304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2836304d26faSStefano Zampini if (!n_D) { 2837304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2838304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2839304d26faSStefano Zampini } 2840304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2841304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2842304d26faSStefano Zampini /* set ksp_D into pcis data */ 2843304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2844304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2845304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2846684f6988SStefano Zampini } 2847304d26faSStefano Zampini 2848304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2849684f6988SStefano Zampini A_RR = 0; 2850684f6988SStefano Zampini if (neumann) { 2851d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 285204708bb6SStefano Zampini PetscInt ibs,mbs; 285304708bb6SStefano Zampini PetscBool issbaij; 285404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2855f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28568ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2857f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2858f4ddd8eeSStefano Zampini PetscInt nn_R; 285981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2860f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2861f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2862f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2863f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2864f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2865f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2866f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2867727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2868f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2869f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2870f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2871f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2872f4ddd8eeSStefano Zampini } 2873f4ddd8eeSStefano Zampini } 2874f4ddd8eeSStefano Zampini /* last check */ 2875d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2876f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2877f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2878f4ddd8eeSStefano Zampini } 2879f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2880f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2881f4ddd8eeSStefano Zampini } 2882a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 2883af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2884af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 288504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 288604708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 288704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 288804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 288904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2890af732b37SStefano Zampini } else { 2891511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 28926816873aSStefano Zampini } 289304708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 289404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 289504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 289604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 289704708bb6SStefano Zampini } else { 2898511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 289904708bb6SStefano Zampini } 290004708bb6SStefano Zampini } 2901a00504b5SStefano Zampini /* extract A_RR */ 2902a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 2903a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2904a00504b5SStefano Zampini 2905a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 290616e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2907a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 290816e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 290916e386b8SStefano Zampini } else { 2910a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 2911a00504b5SStefano Zampini } 2912a00504b5SStefano Zampini } else { 2913a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2914a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 2915a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2916a00504b5SStefano Zampini } 2917a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 2918f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 291916e386b8SStefano Zampini } 29203301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29213301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29226816873aSStefano Zampini } 2923f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2924304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2925304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2926304d26faSStefano Zampini /* default */ 2927304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2928e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2929304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29309577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29319577ea80SStefano Zampini if (issbaij) { 29329577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29339577ea80SStefano Zampini } else { 2934304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29359577ea80SStefano Zampini } 2936304d26faSStefano Zampini /* Allow user's customization */ 2937304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2938304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2939304d26faSStefano Zampini } 2940304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2941304d26faSStefano Zampini if (!n_R) { 2942304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2943304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2944304d26faSStefano Zampini } 29455cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2946df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2947df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2948df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2949d62866d3SStefano Zampini 2950df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2951d62866d3SStefano Zampini } 2952304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2953304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2954684f6988SStefano Zampini } 2955304d26faSStefano Zampini 2956304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29570fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2958684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2959684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29601575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2961684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2962684f6988SStefano Zampini } 2963684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29640fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29650fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29660fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29670fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29680fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2969304d26faSStefano Zampini /* need to be adapted? */ 2970b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2971b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2972b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2973304d26faSStefano Zampini /* print info */ 2974304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2975e604994aSStefano 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); 2976304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2977304d26faSStefano Zampini } 2978b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2979298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2980304d26faSStefano Zampini } 2981684f6988SStefano Zampini } 2982684f6988SStefano Zampini if (neumann) { /* Neumann */ 29830fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 29840fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 29850fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 29860fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 29870fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2988304d26faSStefano Zampini /* need to be adapted? */ 2989b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2990b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2991304d26faSStefano Zampini /* print info */ 2992304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2993e604994aSStefano 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); 2994304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2995304d26faSStefano Zampini } 2996b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2997298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2998304d26faSStefano Zampini } 29990fccc4e9SStefano Zampini } 3000684f6988SStefano Zampini } 30015cbda25cSStefano Zampini /* free Neumann problem's matrix */ 30025cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3003304d26faSStefano Zampini PetscFunctionReturn(0); 3004304d26faSStefano Zampini } 3005304d26faSStefano Zampini 3006304d26faSStefano Zampini #undef __FUNCT__ 3007ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 300880677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3009674ae819SStefano Zampini { 3010674ae819SStefano Zampini PetscErrorCode ierr; 3011674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3012be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3013674ae819SStefano Zampini 3014674ae819SStefano Zampini PetscFunctionBegin; 3015df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 301680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 301720c7b377SStefano Zampini } 301880677318SStefano Zampini if (!pcbddc->switch_static) { 301980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 302080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 302180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 302220c7b377SStefano Zampini } 3023df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 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); 302620c7b377SStefano Zampini } else { 3027df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3028be83ff47SStefano Zampini 3029df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3030df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 303120c7b377SStefano Zampini } 3032be83ff47SStefano Zampini } else { 303380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 303480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 303580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 303680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 303780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 303880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 303980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 304080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3042674ae819SStefano Zampini } 3043674ae819SStefano Zampini } 304472b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 304580677318SStefano Zampini if (applytranspose) { 304680677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 304780677318SStefano Zampini } else { 304880677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 304980677318SStefano Zampini } 3050be83ff47SStefano Zampini } else { 3051df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3052be83ff47SStefano Zampini 3053be83ff47SStefano Zampini if (applytranspose) { 3054df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3055be83ff47SStefano Zampini } else { 3056df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3057be83ff47SStefano Zampini } 3058be83ff47SStefano Zampini } 305980677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 306080677318SStefano Zampini if (!pcbddc->switch_static) { 3061df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 306280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3064be83ff47SStefano Zampini } else { 3065df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3066be83ff47SStefano Zampini 3067df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3068df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3069be83ff47SStefano Zampini } 307080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 307180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 307280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 307380677318SStefano Zampini } 307480677318SStefano Zampini } else { 307580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 307680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 307780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 307880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 307980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 308080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 308180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 308280677318SStefano Zampini } 308380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3087674ae819SStefano Zampini } 3088674ae819SStefano Zampini PetscFunctionReturn(0); 3089674ae819SStefano Zampini } 3090674ae819SStefano Zampini 3091dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3092674ae819SStefano Zampini #undef __FUNCT__ 3093674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3094dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3095674ae819SStefano Zampini { 3096674ae819SStefano Zampini PetscErrorCode ierr; 3097674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3098674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3099674ae819SStefano Zampini const PetscScalar zero = 0.0; 3100674ae819SStefano Zampini 3101674ae819SStefano Zampini PetscFunctionBegin; 3102dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 31034fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3104dc359a40SStefano Zampini if (applytranspose) { 3105674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 31068eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3107dc359a40SStefano Zampini } else { 3108674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3109674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 311015aaf578SStefano Zampini } 31114fee134fSStefano Zampini } else { 31124fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 31134fee134fSStefano Zampini } 3114efc2fbd9SStefano Zampini 3115efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31164f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3117efc2fbd9SStefano Zampini PetscScalar *array; 31184f1b2e48SStefano Zampini PetscInt j; 3119efc2fbd9SStefano Zampini 3120efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31214f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3122efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3123efc2fbd9SStefano Zampini } 3124efc2fbd9SStefano Zampini 312512edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 312612edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 312712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 312812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 312912edc857SStefano Zampini 31309f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 313112edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 313212edc857SStefano Zampini if (pcbddc->coarse_ksp) { 313351694757SStefano Zampini Mat coarse_mat; 3134964fefecSStefano Zampini Vec rhs,sol; 313551694757SStefano Zampini MatNullSpace nullsp; 3136964fefecSStefano Zampini 3137964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3138964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 313951694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 314051694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 314151694757SStefano Zampini if (nullsp) { 314251694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 314351694757SStefano Zampini } 314412edc857SStefano Zampini if (applytranspose) { 3145964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 314612edc857SStefano Zampini } else { 3147964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 314812edc857SStefano Zampini } 314951694757SStefano Zampini if (nullsp) { 315051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 315151694757SStefano Zampini } 315212edc857SStefano Zampini } 3153674ae819SStefano Zampini 3154674ae819SStefano Zampini /* Local solution on R nodes */ 31554fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 315680677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31579f00e9b4SStefano Zampini } 31589f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31599f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 316012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3161674ae819SStefano Zampini 31624fee134fSStefano Zampini /* Sum contributions from the two levels */ 31634fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3164dc359a40SStefano Zampini if (applytranspose) { 3165dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3166dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3167dc359a40SStefano Zampini } else { 3168674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31698eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3170dc359a40SStefano Zampini } 3171efc2fbd9SStefano Zampini /* store p0 */ 31724f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3173efc2fbd9SStefano Zampini PetscScalar *array; 31744f1b2e48SStefano Zampini PetscInt j; 3175efc2fbd9SStefano Zampini 3176efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31774f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3178efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3179efc2fbd9SStefano Zampini } 31804fee134fSStefano Zampini } else { /* expand the coarse solution */ 31814fee134fSStefano Zampini if (applytranspose) { 31824fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31834fee134fSStefano Zampini } else { 31844fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31854fee134fSStefano Zampini } 31864fee134fSStefano Zampini } 3187674ae819SStefano Zampini PetscFunctionReturn(0); 3188674ae819SStefano Zampini } 3189674ae819SStefano Zampini 3190674ae819SStefano Zampini #undef __FUNCT__ 3191674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 319212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3193674ae819SStefano Zampini { 3194674ae819SStefano Zampini PetscErrorCode ierr; 3195674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 319658da7f69SStefano Zampini PetscScalar *array; 319712edc857SStefano Zampini Vec from,to; 3198674ae819SStefano Zampini 3199674ae819SStefano Zampini PetscFunctionBegin; 320012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 320112edc857SStefano Zampini from = pcbddc->coarse_vec; 320212edc857SStefano Zampini to = pcbddc->vec1_P; 320312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 320412edc857SStefano Zampini Vec tvec; 320558da7f69SStefano Zampini 320658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 320758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 320812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 320958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 321058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 321158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 321212edc857SStefano Zampini } 321312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 321412edc857SStefano Zampini from = pcbddc->vec1_P; 321512edc857SStefano Zampini to = pcbddc->coarse_vec; 321612edc857SStefano Zampini } 321712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3218674ae819SStefano Zampini PetscFunctionReturn(0); 3219674ae819SStefano Zampini } 3220674ae819SStefano Zampini 3221674ae819SStefano Zampini #undef __FUNCT__ 3222674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 322312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3224674ae819SStefano Zampini { 3225674ae819SStefano Zampini PetscErrorCode ierr; 3226674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 322758da7f69SStefano Zampini PetscScalar *array; 322812edc857SStefano Zampini Vec from,to; 3229674ae819SStefano Zampini 3230674ae819SStefano Zampini PetscFunctionBegin; 323112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 323212edc857SStefano Zampini from = pcbddc->coarse_vec; 323312edc857SStefano Zampini to = pcbddc->vec1_P; 323412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 323512edc857SStefano Zampini from = pcbddc->vec1_P; 323612edc857SStefano Zampini to = pcbddc->coarse_vec; 323712edc857SStefano Zampini } 323812edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 323912edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 324012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 324112edc857SStefano Zampini Vec tvec; 324258da7f69SStefano Zampini 324312edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 324458da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 324558da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 324658da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 324758da7f69SStefano Zampini } 324858da7f69SStefano Zampini } else { 324958da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 325058da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 325112edc857SStefano Zampini } 325212edc857SStefano Zampini } 3253674ae819SStefano Zampini PetscFunctionReturn(0); 3254674ae819SStefano Zampini } 3255674ae819SStefano Zampini 3256984c4197SStefano Zampini /* uncomment for testing purposes */ 3257984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3258674ae819SStefano Zampini #undef __FUNCT__ 3259674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3260674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3261674ae819SStefano Zampini { 3262674ae819SStefano Zampini PetscErrorCode ierr; 3263674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3264674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3265674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3266984c4197SStefano Zampini /* one and zero */ 3267984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3268984c4197SStefano Zampini /* space to store constraints and their local indices */ 32699162d606SStefano Zampini PetscScalar *constraints_data; 32709162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32719162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32729162d606SStefano Zampini PetscInt *constraints_n; 3273984c4197SStefano Zampini /* iterators */ 3274b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3275984c4197SStefano Zampini /* BLAS integers */ 3276e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3277e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3278c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3279727cdba6SStefano Zampini /* reuse */ 32800e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 32810e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3282984c4197SStefano Zampini /* change of basis */ 3283b3d85658SStefano Zampini PetscBool qr_needed; 32849162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3285984c4197SStefano Zampini /* auxiliary stuff */ 328664efe560SStefano Zampini PetscInt *nnz,*is_indices; 32878a0068c3SStefano Zampini PetscInt ncc; 3288984c4197SStefano Zampini /* some quantities */ 328945a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3290a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3291984c4197SStefano Zampini 3292674ae819SStefano Zampini PetscFunctionBegin; 32938e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 32948e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 32958e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 329616909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3297088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3298088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 32990e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 33000e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 33010e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33020e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33030e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3304088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3305cf5a6209SStefano Zampini 3306cf5a6209SStefano Zampini /* print some info */ 3307cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3308cf5a6209SStefano Zampini IS vertices; 3309cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3310cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3311cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3312cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 33131575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3314cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3315cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3316fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3317fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3318cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33191575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3320cf5a6209SStefano Zampini } 3321cf5a6209SStefano Zampini 3322cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33239162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3324cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3325cf5a6209SStefano Zampini const Vec *nearnullvecs; 3326cf5a6209SStefano Zampini Vec *localnearnullsp; 3327cf5a6209SStefano Zampini PetscScalar *array; 3328cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3329cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3330674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3331b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3332674ae819SStefano Zampini PetscScalar *work; 3333674ae819SStefano Zampini PetscReal *singular_vals; 3334674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3335674ae819SStefano Zampini PetscReal *rwork; 3336674ae819SStefano Zampini #endif 3337674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3338674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3339674ae819SStefano Zampini #else 3340964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3341964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3342674ae819SStefano Zampini #endif 3343674ae819SStefano Zampini 3344674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3345d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3346d06fc5fdSStefano Zampini /* free unneeded index sets */ 3347d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3348d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3349674ae819SStefano Zampini } 3350d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3351d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3352d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3353d06fc5fdSStefano Zampini } 3354d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3355d06fc5fdSStefano Zampini n_ISForEdges = 0; 3356d06fc5fdSStefano Zampini } 3357d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3358d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3359d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3360d06fc5fdSStefano Zampini } 3361d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3362d06fc5fdSStefano Zampini n_ISForFaces = 0; 3363d06fc5fdSStefano Zampini } 336470022509SStefano Zampini 336570022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 336670022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 336770022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 336870022509SStefano Zampini if (pcbddc->NullSpace) { 336970022509SStefano Zampini PetscBool tbool[2],gbool[2]; 337070022509SStefano Zampini 337170022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3372b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3373d06fc5fdSStefano Zampini if (!ISForEdges) { 3374d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3375d06fc5fdSStefano Zampini } 3376b8ffe317SStefano Zampini } 3377d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3378d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3379b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3380d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3381d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 338298a51de6SStefano Zampini } 338370022509SStefano Zampini #endif 338408122e43SStefano Zampini 3385674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3386674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3387674ae819SStefano Zampini if (nearnullsp) { 3388674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3389f4ddd8eeSStefano Zampini /* remove any stored info */ 3390f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3391f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3392f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3393f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3394f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3395473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3396f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3397f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3398f4ddd8eeSStefano Zampini } 3399984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3400984c4197SStefano Zampini nnsp_size = 0; 3401674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3402674ae819SStefano Zampini } 3403984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3404984c4197SStefano Zampini max_constraints = nnsp_size; 3405984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3406984c4197SStefano Zampini 3407674ae819SStefano Zampini /* 3408674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 34099162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 34109162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 34119162d606SStefano Zampini There can be multiple constraints per connected component 3412674ae819SStefano Zampini */ 3413674ae819SStefano Zampini n_vertices = 0; 3414674ae819SStefano Zampini if (ISForVertices) { 3415674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3416674ae819SStefano Zampini } 34179162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34189162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34199162d606SStefano Zampini 34209162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34219162d606SStefano Zampini total_counts *= max_constraints; 3422674ae819SStefano Zampini total_counts += n_vertices; 34234641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34249162d606SStefano Zampini 3425674ae819SStefano Zampini total_counts = 0; 3426674ae819SStefano Zampini max_size_of_constraint = 0; 3427674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34289162d606SStefano Zampini IS used_is; 3429674ae819SStefano Zampini if (i<n_ISForEdges) { 34309162d606SStefano Zampini used_is = ISForEdges[i]; 3431674ae819SStefano Zampini } else { 34329162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3433674ae819SStefano Zampini } 34349162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3435674ae819SStefano Zampini total_counts += j; 3436674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3437674ae819SStefano Zampini } 34389162d606SStefano 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); 34399162d606SStefano Zampini 3440984c4197SStefano Zampini /* get local part of global near null space vectors */ 3441785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3442984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3443984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3444e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3445e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3446984c4197SStefano Zampini } 3447674ae819SStefano Zampini 3448242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3449242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3450a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3451242a89d7SStefano Zampini 3452984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3453a773dcb8SStefano Zampini if (!skip_lapack) { 3454674ae819SStefano Zampini PetscScalar temp_work; 3455911cabfeSStefano Zampini 3456674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3457984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3458785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3459785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3460785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3461674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3462785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3463674ae819SStefano Zampini #endif 3464674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3465c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3466c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3467674ae819SStefano Zampini lwork = -1; 3468674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3469674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3470c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3471674ae819SStefano Zampini #else 3472c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3473674ae819SStefano Zampini #endif 3474674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3475984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3476674ae819SStefano Zampini #else /* on missing GESVD */ 3477674ae819SStefano Zampini /* SVD */ 3478674ae819SStefano Zampini PetscInt max_n,min_n; 3479674ae819SStefano Zampini max_n = max_size_of_constraint; 3480984c4197SStefano Zampini min_n = max_constraints; 3481984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3482674ae819SStefano Zampini min_n = max_size_of_constraint; 3483984c4197SStefano Zampini max_n = max_constraints; 3484674ae819SStefano Zampini } 3485785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3486674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3487785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3488674ae819SStefano Zampini #endif 3489674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3490674ae819SStefano Zampini lwork = -1; 3491e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3492e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3493b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3494674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3495674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 34969162d606SStefano 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)); 3497674ae819SStefano Zampini #else 34989162d606SStefano 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)); 3499674ae819SStefano Zampini #endif 3500674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3501984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3502984c4197SStefano Zampini #endif /* on missing GESVD */ 3503674ae819SStefano Zampini /* Allocate optimal workspace */ 3504674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3505854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3506674ae819SStefano Zampini } 3507674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3508674ae819SStefano Zampini total_counts = 0; 35099162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 35109162d606SStefano Zampini constraints_data_ptr[0] = 0; 3511674ae819SStefano Zampini /* vertices */ 35129162d606SStefano Zampini if (n_vertices) { 3513674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 35149162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3515674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35169162d606SStefano Zampini constraints_n[total_counts] = 1; 35179162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35189162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35199162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3520674ae819SStefano Zampini total_counts++; 3521674ae819SStefano Zampini } 3522674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3523674ae819SStefano Zampini n_vertices = total_counts; 3524674ae819SStefano Zampini } 3525984c4197SStefano Zampini 3526674ae819SStefano Zampini /* edges and faces */ 35279162d606SStefano Zampini total_counts_cc = total_counts; 3528911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35299162d606SStefano Zampini IS used_is; 35309162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35319162d606SStefano Zampini 3532911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35339162d606SStefano Zampini used_is = ISForEdges[ncc]; 3534984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3535674ae819SStefano Zampini } else { 35369162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3537984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3538674ae819SStefano Zampini } 3539674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35409162d606SStefano Zampini 35419162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35429162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3543984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3544984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3545674ae819SStefano Zampini if (nnsp_has_cnst) { 35465b08dc53SStefano Zampini PetscScalar quad_value; 35479162d606SStefano Zampini 35489162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35499162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35509162d606SStefano Zampini 3551a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3552674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3553a773dcb8SStefano Zampini } else { 3554a773dcb8SStefano Zampini quad_value = 1.0; 3555a773dcb8SStefano Zampini } 3556674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35579162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3558674ae819SStefano Zampini } 35599162d606SStefano Zampini temp_constraints++; 3560674ae819SStefano Zampini total_counts++; 3561674ae819SStefano Zampini } 3562674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3563984c4197SStefano Zampini PetscReal real_value; 35649162d606SStefano Zampini PetscScalar *ptr_to_data; 35659162d606SStefano Zampini 3566984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35679162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3568674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35699162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3570674ae819SStefano Zampini } 3571984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3572984c4197SStefano Zampini /* check if array is null on the connected component */ 3573e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35749162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35755b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3576674ae819SStefano Zampini temp_constraints++; 3577674ae819SStefano Zampini total_counts++; 35789162d606SStefano Zampini if (!idxs_copied) { 35799162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35809162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3581674ae819SStefano Zampini } 3582674ae819SStefano Zampini } 35839162d606SStefano Zampini } 35849162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 358545a1bb75SStefano Zampini valid_constraints = temp_constraints; 3586eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3587a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 35889162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 35899162d606SStefano Zampini 35909162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3591a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35929162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3593a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 35949162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3595a773dcb8SStefano Zampini } else { /* perform SVD */ 3596984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 35979162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3598674ae819SStefano Zampini 3599674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3600984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3601984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3602984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3603984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3604984c4197SStefano Zampini from that computed using LAPACKgesvd 3605984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3606984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3607984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3608674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3609e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3610984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3611674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3612674ae819SStefano Zampini for (k=0;k<j+1;k++) { 36139162d606SStefano 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)); 3614674ae819SStefano Zampini } 3615674ae819SStefano Zampini } 3616e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3617e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3618e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3619674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3620c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3621674ae819SStefano Zampini #else 3622c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3623674ae819SStefano Zampini #endif 3624674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3625984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3626984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3627674ae819SStefano Zampini j = 0; 3628984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3629674ae819SStefano Zampini total_counts = total_counts-j; 363045a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3631e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3632c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3633c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3634c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3635c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3636c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3637c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3638674ae819SStefano Zampini if (j<temp_constraints) { 3639984c4197SStefano Zampini PetscInt ii; 3640984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3641674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36429162d606SStefano 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)); 3643674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3644984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3645674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36469162d606SStefano 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]; 3647674ae819SStefano Zampini } 3648674ae819SStefano Zampini } 3649674ae819SStefano Zampini } 3650674ae819SStefano Zampini #else /* on missing GESVD */ 3651e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3652e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3653b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3654674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3655674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36569162d606SStefano 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)); 3657674ae819SStefano Zampini #else 36589162d606SStefano 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)); 3659674ae819SStefano Zampini #endif 3660984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3661674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3662984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3663e310c8b4SStefano Zampini k = temp_constraints; 3664e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3665674ae819SStefano Zampini j = 0; 3666e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 366745a1bb75SStefano Zampini valid_constraints = k-j; 3668911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3669984c4197SStefano Zampini #endif /* on missing GESVD */ 3670674ae819SStefano Zampini } 3671a773dcb8SStefano Zampini } 36729162d606SStefano Zampini /* update pointers information */ 36739162d606SStefano Zampini if (valid_constraints) { 36749162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36759162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36769162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36779162d606SStefano Zampini /* set change_of_basis flag */ 367845a1bb75SStefano Zampini if (boolforchange) { 3679b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 36809162d606SStefano Zampini } 3681b3d85658SStefano Zampini total_counts_cc++; 368245a1bb75SStefano Zampini } 368345a1bb75SStefano Zampini } 3684984c4197SStefano Zampini /* free workspace */ 36858f1c130eSStefano Zampini if (!skip_lapack) { 3686984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3687984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3688984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3689984c4197SStefano Zampini #endif 3690984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3691984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3692984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3693984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3694984c4197SStefano Zampini #endif 3695984c4197SStefano Zampini } 3696984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3697984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3698984c4197SStefano Zampini } 3699984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3700cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3701cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3702cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3703cf5a6209SStefano Zampini } 3704cf5a6209SStefano Zampini if (n_ISForFaces) { 3705cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3706cf5a6209SStefano Zampini } 3707cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3708cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3709cf5a6209SStefano Zampini } 3710cf5a6209SStefano Zampini if (n_ISForEdges) { 3711cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3712cf5a6209SStefano Zampini } 3713cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 371408122e43SStefano Zampini } else { 371508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3716984c4197SStefano Zampini 371708122e43SStefano Zampini total_counts = 0; 371808122e43SStefano Zampini n_vertices = 0; 3719d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3720d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 372108122e43SStefano Zampini } 372208122e43SStefano Zampini max_constraints = 0; 37239162d606SStefano Zampini total_counts_cc = 0; 372408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 372508122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37269162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 372708122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 372808122e43SStefano Zampini } 37299162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37309162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37319162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37329162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 373374d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37349162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37359162d606SStefano Zampini total_counts_cc = 0; 37369162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37379162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37389162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 373908122e43SStefano Zampini } 374008122e43SStefano Zampini } 37419162d606SStefano Zampini #if 0 37429162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37439162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37449162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37459162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37469162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37479162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37489162d606SStefano Zampini } 37499162d606SStefano Zampini printf("\n"); 37509162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37519162d606SStefano Zampini } 37521b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37538bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37541b968477SStefano Zampini } 37551b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37568bec7fa6SStefano 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]); 37571b968477SStefano Zampini } 375808122e43SStefano Zampini #endif 375908122e43SStefano Zampini 37608bec7fa6SStefano Zampini max_size_of_constraint = 0; 37619162d606SStefano 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]); 37629162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 376308122e43SStefano Zampini /* Change of basis */ 3764b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 376508122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 376608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 376708122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3768b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 376908122e43SStefano Zampini } 377008122e43SStefano Zampini } 377108122e43SStefano Zampini } 377208122e43SStefano Zampini } 3773984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37744f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 377508122e43SStefano Zampini 37769162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37779162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 37786c4ed002SBarry 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); 3779674ae819SStefano Zampini 3780674ae819SStefano Zampini /* Create constraint matrix */ 3781674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 378216f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3783984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3784984c4197SStefano Zampini 3785984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3786a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3787a717540cSStefano Zampini qr_needed = PETSC_FALSE; 378874d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3789984c4197SStefano Zampini total_primal_vertices=0; 3790b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 37919162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37929162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 379372b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 37949162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3795b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 379664efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 37979162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37989162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3799a717540cSStefano Zampini } 3800b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 380174d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3802a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3803a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3804a717540cSStefano Zampini } 3805fa434743SStefano Zampini } else { 3806b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3807fa434743SStefano Zampini } 3808a717540cSStefano Zampini } 3809b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3810b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3811674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 381270022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3813b3d85658SStefano Zampini 38144f1b2e48SStefano 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); 38150e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38160e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3817984c4197SStefano Zampini 3818984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 381974d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3820785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3821984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 382274d5cdf7SStefano Zampini 3823984c4197SStefano Zampini j = total_primal_vertices; 382474d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3825b3d85658SStefano Zampini cum = total_primal_vertices; 38269162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38274641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3828b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3829b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3830b3d85658SStefano Zampini cum++; 38319162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 383274d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 383374d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 383474d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 383574d5cdf7SStefano Zampini } 38369162d606SStefano Zampini j += constraints_n[i]; 3837674ae819SStefano Zampini } 3838674ae819SStefano Zampini } 3839674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3840674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3841088faed8SStefano Zampini 3842674ae819SStefano Zampini /* set values in constraint matrix */ 3843984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38440e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3845674ae819SStefano Zampini } 3846984c4197SStefano Zampini total_counts = total_primal_vertices; 38479162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38484641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38499162d606SStefano Zampini PetscInt *cols; 38509162d606SStefano Zampini 38519162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38529162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38539162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38549162d606SStefano Zampini PetscInt row = total_counts+k; 38559162d606SStefano Zampini PetscScalar *vals; 38569162d606SStefano Zampini 38579162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38589162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38599162d606SStefano Zampini } 38609162d606SStefano Zampini total_counts += constraints_n[i]; 3861674ae819SStefano Zampini } 3862674ae819SStefano Zampini } 3863674ae819SStefano Zampini /* assembling */ 3864674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3865674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3866088faed8SStefano Zampini 3867984c4197SStefano Zampini /* 38686a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3869984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3870f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 3871984c4197SStefano Zampini */ 3872674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3873674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3874026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3875984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3876984c4197SStefano Zampini /* working stuff for GEQRF */ 387781d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3878984c4197SStefano Zampini PetscBLASInt lqr_work; 3879984c4197SStefano Zampini /* working stuff for UNGQR */ 3880984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3881984c4197SStefano Zampini PetscBLASInt lgqr_work; 3882984c4197SStefano Zampini /* working stuff for TRTRS */ 3883984c4197SStefano Zampini PetscScalar *trs_rhs; 38843f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3885984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3886984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3887984c4197SStefano Zampini PetscScalar *start_vals; 3888984c4197SStefano Zampini /* working stuff for values insertion */ 38894641a718SStefano Zampini PetscBT is_primal; 389064efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3891906d46d4SStefano Zampini /* matrix sizes */ 3892906d46d4SStefano Zampini PetscInt global_size,local_size; 3893906d46d4SStefano Zampini /* temporary change of basis */ 3894906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3895cf5a6209SStefano Zampini /* extra space for debugging */ 3896cf5a6209SStefano Zampini PetscScalar *dbg_work; 3897984c4197SStefano Zampini 3898906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3899906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 390016f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3901bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3902906d46d4SStefano Zampini /* nonzeros for local mat */ 3903bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 39041dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3905bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 39061dd7afcfSStefano Zampini } else { 39071dd7afcfSStefano Zampini const PetscInt *ii; 39081dd7afcfSStefano Zampini PetscInt n; 39091dd7afcfSStefano Zampini PetscBool flg_row; 39101dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39111dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 39121dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39131dd7afcfSStefano Zampini } 39149162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3915a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3917a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39189162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3919a717540cSStefano Zampini } else { 39209162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39219162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3922a717540cSStefano Zampini } 3923a717540cSStefano Zampini } 3924a717540cSStefano Zampini } 3925906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3926bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 39271dd7afcfSStefano Zampini /* Set interior change in the matrix */ 39281dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3929bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3930906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3931a717540cSStefano Zampini } 39321dd7afcfSStefano Zampini } else { 39331dd7afcfSStefano Zampini const PetscInt *ii,*jj; 39341dd7afcfSStefano Zampini PetscScalar *aa; 39351dd7afcfSStefano Zampini PetscInt n; 39361dd7afcfSStefano Zampini PetscBool flg_row; 39371dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39381dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39391dd7afcfSStefano Zampini for (i=0;i<n;i++) { 39401dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 39411dd7afcfSStefano Zampini } 39421dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39431dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39441dd7afcfSStefano Zampini } 3945a717540cSStefano Zampini 3946a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3947a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3948a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3949a717540cSStefano Zampini } 3950a717540cSStefano Zampini 3951a717540cSStefano Zampini 3952a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3953a717540cSStefano Zampini /* 3954a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3955a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3956a717540cSStefano Zampini 3957a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3958a717540cSStefano Zampini 3959a6b551f4SStefano 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) 3960a6b551f4SStefano Zampini 3961a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3962a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3963a717540cSStefano Zampini | ... | 3964a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3965a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3966a717540cSStefano Zampini 3967a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3968a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3969a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3970a6b551f4SStefano Zampini 3971a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3972a717540cSStefano Zampini */ 3973a717540cSStefano Zampini if (qr_needed) { 3974984c4197SStefano Zampini /* space to store Q */ 3975854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3976984c4197SStefano Zampini /* first we issue queries for optimal work */ 39773f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39783f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 39793f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3980984c4197SStefano Zampini lqr_work = -1; 39813f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3982984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3983984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3984785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3985984c4197SStefano Zampini lgqr_work = -1; 39863f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39873f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 39883f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 39893f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39903f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 39913f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3992984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3993984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3994785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3995984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3996785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3997984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3998785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3999a717540cSStefano Zampini /* allocating workspace for check */ 4000a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4001cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4002a717540cSStefano Zampini } 4003a717540cSStefano Zampini } 4004984c4197SStefano Zampini /* array to store whether a node is primal or not */ 40054641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4006473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 40070e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 40086c4ed002SBarry 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); 400939e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 401039e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 401139e2fb2aSStefano Zampini } 401239e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4013984c4197SStefano Zampini 4014a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 40159162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40174641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4018984c4197SStefano Zampini /* get constraint info */ 40199162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4020984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4021984c4197SStefano Zampini 4022984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40239162d606SStefano 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); 4024674ae819SStefano Zampini } 4025984c4197SStefano Zampini 4026fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4027a717540cSStefano Zampini 4028a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4029a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40309162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4031a717540cSStefano Zampini } 4032984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40339162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4034984c4197SStefano Zampini 4035984c4197SStefano Zampini /* compute QR decomposition of constraints */ 40363f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40373f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40383f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4039674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40403f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4041984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4042674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4043984c4197SStefano Zampini 4044984c4197SStefano Zampini /* explictly compute R^-T */ 4045984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4046984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40473f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40483f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40493f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40503f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4051984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40523f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4053984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4054984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4055984c4197SStefano Zampini 4056a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40573f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40593f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4061984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40623f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4063984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4064984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4065984c4197SStefano Zampini 4066984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4067984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4068984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40693f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40703f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40713f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40723f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40733f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4075984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40769162d606SStefano 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)); 4077984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40789162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4079984c4197SStefano Zampini 4080984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 40819162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4082984c4197SStefano Zampini /* insert cols for primal dofs */ 4083984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4084984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 40859162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4086906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4087984c4197SStefano Zampini } 4088984c4197SStefano Zampini /* insert cols for dual dofs */ 4089984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 40909162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4091984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 40929162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4093906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4094984c4197SStefano Zampini j++; 4095674ae819SStefano Zampini } 4096674ae819SStefano Zampini } 4097984c4197SStefano Zampini 4098984c4197SStefano Zampini /* check change of basis */ 4099984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4100984c4197SStefano Zampini PetscInt ii,jj; 4101984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4102c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4103c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4104c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4105c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4106c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4107c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4108984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4109cf5a6209SStefano 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)); 4110984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4111984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4112984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4113cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4114cf5a6209SStefano 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; 4115674ae819SStefano Zampini } 4116674ae819SStefano Zampini } 4117984c4197SStefano Zampini if (!valid_qr) { 411822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4119984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4120984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4121cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4122cf5a6209SStefano 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])); 4123674ae819SStefano Zampini } 4124cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4125cf5a6209SStefano 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])); 4126984c4197SStefano Zampini } 4127984c4197SStefano Zampini } 4128984c4197SStefano Zampini } 4129674ae819SStefano Zampini } else { 413022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4131674ae819SStefano Zampini } 4132674ae819SStefano Zampini } 4133a717540cSStefano Zampini } else { /* simple transformation block */ 4134a717540cSStefano Zampini PetscInt row,col; 4135a6b551f4SStefano Zampini PetscScalar val,norm; 4136a6b551f4SStefano Zampini 4137a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41389162d606SStefano 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)); 4139a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 41409162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 41419162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4142bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 41439162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4144906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41459162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4146a717540cSStefano Zampini } else { 4147a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41489162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4149a717540cSStefano Zampini if (row != col) { 41509162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4151a717540cSStefano Zampini } else { 41529162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4153a717540cSStefano Zampini } 4154906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4155a717540cSStefano Zampini } 4156a717540cSStefano Zampini } 4157a717540cSStefano Zampini } 415898a51de6SStefano Zampini if (pcbddc->dbg_flag) { 415922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4160a717540cSStefano Zampini } 4161674ae819SStefano Zampini } 4162984c4197SStefano Zampini } else { 4163984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41649162d606SStefano 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); 4165674ae819SStefano Zampini } 4166674ae819SStefano Zampini } 4167674ae819SStefano Zampini } 4168a717540cSStefano Zampini 4169a717540cSStefano Zampini /* free workspace */ 4170a717540cSStefano Zampini if (qr_needed) { 4171984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4172cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4173984c4197SStefano Zampini } 4174984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4175984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4176984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4177984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4178984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4179674ae819SStefano Zampini } 4180a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4181906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4182906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4183906d46d4SStefano Zampini 4184906d46d4SStefano Zampini /* assembling of global change of variable */ 418588c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4186bbb9e6c6SStefano Zampini Mat tmat; 418716f15bc4SStefano Zampini PetscInt bs; 418816f15bc4SStefano Zampini 4189906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4190906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4191bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4192bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4193bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4194bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 419516f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 419616f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4197906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4198bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4199bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4200bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4201bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4202bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4203e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4204e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4205bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4206bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 420788c03ad3SStefano Zampini 4208906d46d4SStefano Zampini /* check */ 4209906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4210906d46d4SStefano Zampini PetscReal error; 4211906d46d4SStefano Zampini Vec x,x_change; 4212906d46d4SStefano Zampini 4213906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4214906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4215906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4216906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4217e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4218e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4219bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4220e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4221e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4222906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4223906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4224906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4225906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4226bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4227906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4228906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4229906d46d4SStefano Zampini } 4230b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4231b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4232b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4233bf3a8328SStefano Zampini 4234bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4235bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4236bf3a8328SStefano Zampini } 4237b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4238ac632422SStefano Zampini Mat S_new,tmat; 4239bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4240bbb9e6c6SStefano Zampini 4241bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 42426816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4243bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4244bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4245bf3a8328SStefano Zampini IS is_V; 4246b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4247b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4248b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4249b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4250b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4251bf3a8328SStefano Zampini } 4252bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4253ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4254b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4255ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4256bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4257bf3a8328SStefano Zampini const PetscScalar *array; 4258bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4259bf3a8328SStefano Zampini PetscInt i,n_V; 4260bf3a8328SStefano Zampini 4261b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4262b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4263b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4264b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4265b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4266b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4267b087196eSStefano Zampini PetscScalar val; 4268b087196eSStefano Zampini PetscInt idx; 4269b087196eSStefano Zampini 4270b087196eSStefano Zampini idx = idxs_V[i]; 4271b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4272b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4273b087196eSStefano Zampini } 4274b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4275b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4276bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4277bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4278bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4279bf3a8328SStefano Zampini } 4280ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4281ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4282ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4283ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4284b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4285ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4286bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4287b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4288bf3a8328SStefano Zampini } 4289ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4290ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4291ac632422SStefano Zampini } 4292b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 429388c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4294b96c3477SStefano Zampini } 4295c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4296c9db6a07SStefano Zampini if (sub_schurs->change) { 4297c9db6a07SStefano Zampini PetscInt i; 4298c9db6a07SStefano Zampini 4299c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4300c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4301c9db6a07SStefano Zampini } 4302c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4303c9db6a07SStefano Zampini } 4304b96c3477SStefano Zampini } 430516909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 430616909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 430716909a7fSStefano Zampini } else { 4308906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 430916909a7fSStefano Zampini } 43101dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 43111dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43121dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43131dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->change_interior,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43141dd7afcfSStefano Zampini } 431572b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 431672b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 431772b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 431872b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 431972b8c272SStefano Zampini } 43201dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 43211dd7afcfSStefano Zampini PetscBool needglobal = PETSC_FALSE; 43221dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43231dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43241dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&needglobal,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43251dd7afcfSStefano Zampini } 43261dd7afcfSStefano Zampini if (!needglobal && pcbddc->user_ChangeOfBasisMatrix) { 4327b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4328b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4329906d46d4SStefano Zampini } else { 43301dd7afcfSStefano Zampini Mat benign_global = NULL; 43311dd7afcfSStefano Zampini if (needglobal) { 43321dd7afcfSStefano Zampini Mat tmat; 43331dd7afcfSStefano Zampini 43341dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 43351dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 43361dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 43371dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43381dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43391dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 43401dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43411dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43421dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 43431dd7afcfSStefano Zampini if (pcbddc->benign_change) { 43441dd7afcfSStefano Zampini Mat M; 43451dd7afcfSStefano Zampini 43461dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 43471dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 43481dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 43491dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4350906d46d4SStefano Zampini } else { 43511dd7afcfSStefano Zampini Mat eye; 43521dd7afcfSStefano Zampini PetscScalar *array; 43531dd7afcfSStefano Zampini 43541dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43551dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 43561dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 43571dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4358906d46d4SStefano Zampini } 43591dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43601dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43611dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43621dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 43631dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 43641dd7afcfSStefano Zampini } 43651dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 43661dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 43671dd7afcfSStefano Zampini } 43681dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 43691dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 43701dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 43711dd7afcfSStefano Zampini } else if (needglobal) { 43721dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 43731dd7afcfSStefano Zampini } 43741dd7afcfSStefano Zampini } 437516909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 437616909a7fSStefano Zampini IS is_global; 437716909a7fSStefano Zampini const PetscInt *gidxs; 437816909a7fSStefano Zampini 437916909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 438016909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 438116909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 438216909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 438316909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 438416909a7fSStefano Zampini } 43851dd7afcfSStefano Zampini } 43861dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 43871dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4388b9b85e73SStefano Zampini } 4389a717540cSStefano Zampini 439072b8c272SStefano Zampini if (!pcbddc->fake_change) { 43914f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 43924f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 43934f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 43944f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4395019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4396019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4397019a44ceSStefano Zampini pcbddc->local_primal_size++; 4398019a44ceSStefano Zampini } 4399019a44ceSStefano Zampini 4400019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4401727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4402727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 44039f47a83aSStefano 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); 4404c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 44050e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 44069f47a83aSStefano 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); 4407727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4408727cdba6SStefano Zampini } 44090e6343abSStefano Zampini } 4410727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4411b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 441272b8c272SStefano Zampini } 441372b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4414727cdba6SStefano Zampini 4415a717540cSStefano Zampini /* flush dbg viewer */ 4416b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4417b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4418b8ffe317SStefano Zampini } 4419a717540cSStefano Zampini 4420e310c8b4SStefano Zampini /* free workspace */ 4421a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 44224641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 442308122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 44249162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 44259162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 442608122e43SStefano Zampini } else { 44279162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 44289162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 44299162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 443008122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 443108122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 44329162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 44339162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 443408122e43SStefano Zampini } 4435674ae819SStefano Zampini PetscFunctionReturn(0); 4436674ae819SStefano Zampini } 4437674ae819SStefano Zampini 4438674ae819SStefano Zampini #undef __FUNCT__ 4439674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4440674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4441674ae819SStefano Zampini { 4442674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4443674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4444674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 444514f95afaSStefano Zampini PetscInt ierr,i,N; 4446674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4447674ae819SStefano Zampini 4448674ae819SStefano Zampini PetscFunctionBegin; 44498e61c736SStefano Zampini /* Reset previously computed graph */ 44508e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4451674ae819SStefano Zampini /* Init local Graph struct */ 44527fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 44533bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4454674ae819SStefano Zampini 4455575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 44565099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 44575099eff2SStefano 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); 4458575ad6abSStefano Zampini } 44599577ea80SStefano Zampini 4460674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4461d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 44624d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 44634d379d7bSStefano Zampini PetscInt nvtxs; 4464e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4465674ae819SStefano Zampini 44662fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 44672fffb893SStefano Zampini if (flg_row) { 44684d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4469b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 44702fffb893SStefano Zampini } 44712fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4472674ae819SStefano Zampini } 44739b28b941SStefano Zampini if (pcbddc->dbg_flag) { 44749b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4475674ae819SStefano Zampini } 4476674ae819SStefano Zampini 447763602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 447863602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 447963602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 448095ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 448163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4482e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 448363602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4484674ae819SStefano Zampini } 448563602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 448663602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 448763602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4488674ae819SStefano Zampini } 448963602bcaSStefano Zampini } else { 449063602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 449163602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4492854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 449363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 449463602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 449563602bcaSStefano Zampini } 449663602bcaSStefano Zampini } 4497674ae819SStefano Zampini } 4498674ae819SStefano Zampini 4499674ae819SStefano Zampini /* Setup of Graph */ 4500785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4501e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4502785d1243SStefano Zampini } 4503785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4504e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4505785d1243SStefano Zampini } 450630368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 450730368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 450830368db7SStefano Zampini } 450914f95afaSStefano 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); 4510674ae819SStefano Zampini 45114f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 45124f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 45134f1b2e48SStefano Zampini PetscInt *local_subs; 45144f1b2e48SStefano Zampini 45154f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 45164f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 45174f1b2e48SStefano Zampini const PetscInt *idxs; 45184f1b2e48SStefano Zampini PetscInt nl,j; 45194f1b2e48SStefano Zampini 45204f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 45214f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45224f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 45234f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 45244f1b2e48SStefano Zampini } 45254f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45264f1b2e48SStefano Zampini } 45274f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 45284f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 45294f1b2e48SStefano Zampini } 45304f1b2e48SStefano Zampini 4531674ae819SStefano Zampini /* Graph's connected components analysis */ 4532674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4533674ae819SStefano Zampini 4534674ae819SStefano Zampini /* print some info to stdout */ 4535674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4536302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4537674ae819SStefano Zampini } 4538674ae819SStefano Zampini PetscFunctionReturn(0); 4539674ae819SStefano Zampini } 4540674ae819SStefano Zampini 4541dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4542674ae819SStefano Zampini #undef __FUNCT__ 4543674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4544dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4545674ae819SStefano Zampini { 4546dc456d91SStefano Zampini PetscSF sf; 4547dc456d91SStefano Zampini PetscLayout map; 4548dc456d91SStefano Zampini const PetscInt *idxs; 4549dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4550dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4551dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4552dc456d91SStefano Zampini PetscMPIInt commsize; 4553674ae819SStefano Zampini PetscBool first_found; 4554674ae819SStefano Zampini PetscErrorCode ierr; 4555674ae819SStefano Zampini 4556674ae819SStefano Zampini PetscFunctionBegin; 4557dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4558dc456d91SStefano Zampini if (subset_mult) { 4559dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4560dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4561dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4562674ae819SStefano Zampini } 4563dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4564dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4565dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4566dc456d91SStefano Zampini for (i=0;i<n;i++) { 4567dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4568dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4569674ae819SStefano Zampini } 4570dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4571b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4572dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4573dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4574dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4575dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4576dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4577dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4578dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4579dc456d91SStefano Zampini 4580dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4581dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4582dc456d91SStefano Zampini if (subset_mult) { 4583dc456d91SStefano Zampini const PetscInt* idxs_mult; 4584dc456d91SStefano Zampini 4585dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4586dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4587dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4588674ae819SStefano Zampini } else { 4589dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4590674ae819SStefano Zampini } 4591dc456d91SStefano Zampini /* local size of new subset */ 4592dc456d91SStefano Zampini n_n = 0; 4593dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4594dc456d91SStefano Zampini 4595dc456d91SStefano Zampini /* global indexes in layout */ 4596dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4597dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4598dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4599dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4600dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4601dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4602dc456d91SStefano Zampini 4603dc456d91SStefano Zampini /* reduce from leaves to roots */ 4604dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 460564a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 460664a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4607dc456d91SStefano Zampini 4608dc456d91SStefano Zampini /* count indexes in local part of layout */ 4609674ae819SStefano Zampini nlocals = 0; 4610674ae819SStefano Zampini first_index = -1; 4611674ae819SStefano Zampini first_found = PETSC_FALSE; 4612dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4613dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4614674ae819SStefano Zampini first_found = PETSC_TRUE; 4615674ae819SStefano Zampini first_index = i; 4616674ae819SStefano Zampini } 4617dc456d91SStefano Zampini nlocals += root_data[i]; 4618674ae819SStefano Zampini } 4619dc456d91SStefano Zampini 4620dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 46215fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4622dc456d91SStefano Zampini start = 0; 462364a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46245fa240b1SStefano Zampini #else 462564a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46265fa240b1SStefano Zampini start = start-nlocals; 46275fa240b1SStefano Zampini #endif 46285fa240b1SStefano Zampini 4629dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4630dc456d91SStefano Zampini *N_n = start + nlocals; 4631dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4632dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4633674ae819SStefano Zampini } 46345fa240b1SStefano Zampini 46355fa240b1SStefano Zampini /* adapt root data with cumulative */ 4636674ae819SStefano Zampini if (first_found) { 4637dc456d91SStefano Zampini PetscInt old_index; 4638dc456d91SStefano Zampini 4639dc456d91SStefano Zampini root_data[first_index] += start; 4640674ae819SStefano Zampini old_index = first_index; 4641dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4642dc456d91SStefano Zampini if (root_data[i]) { 4643dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4644674ae819SStefano Zampini old_index = i; 4645674ae819SStefano Zampini } 4646674ae819SStefano Zampini } 4647674ae819SStefano Zampini } 4648dc456d91SStefano Zampini 4649dc456d91SStefano Zampini /* from roots to leaves */ 4650dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4651dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4652dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4653dc456d91SStefano Zampini 4654dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4655dc456d91SStefano Zampini if (subset_mult) { 4656dc456d91SStefano Zampini const PetscInt* idxs_mult; 4657dc456d91SStefano Zampini PetscInt cum; 4658dc456d91SStefano Zampini 4659dc456d91SStefano Zampini cum = 0; 4660dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4661dc456d91SStefano Zampini for (i=0;i<n;i++) { 4662dc456d91SStefano Zampini PetscInt j; 4663dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4664674ae819SStefano Zampini } 4665dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4666674ae819SStefano Zampini } else { 4667dc456d91SStefano Zampini for (i=0;i<n;i++) { 4668dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4669674ae819SStefano Zampini } 4670674ae819SStefano Zampini } 4671dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4672dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4673674ae819SStefano Zampini PetscFunctionReturn(0); 4674674ae819SStefano Zampini } 46759a7d3425SStefano Zampini 46769a7d3425SStefano Zampini #undef __FUNCT__ 46779a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 46789a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 46799a7d3425SStefano Zampini { 46809a7d3425SStefano Zampini PetscInt i,j; 46819a7d3425SStefano Zampini PetscScalar *alphas; 46829a7d3425SStefano Zampini PetscErrorCode ierr; 46839a7d3425SStefano Zampini 46849a7d3425SStefano Zampini PetscFunctionBegin; 46859a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4686785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 46879a7d3425SStefano Zampini for (i=0;i<n;i++) { 46889a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 46899a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 46909a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 46919a7d3425SStefano Zampini } 46929a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 46939a7d3425SStefano Zampini PetscFunctionReturn(0); 46949a7d3425SStefano Zampini } 46959a7d3425SStefano Zampini 4696e7931f94SStefano Zampini #undef __FUNCT__ 469770cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4698b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4699e7931f94SStefano Zampini { 470052e5ac9dSStefano Zampini IS ranks_send_to; 4701e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4702e7931f94SStefano Zampini PetscMPIInt size,rank,color; 470352e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 470452e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 47053837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 47062b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4707e7931f94SStefano Zampini PetscSubcomm subcomm; 470852e5ac9dSStefano Zampini PetscErrorCode ierr; 4709a57a6d2fSStefano Zampini 4710e7931f94SStefano Zampini PetscFunctionBegin; 4711c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 4712c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4713c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4714e7931f94SStefano Zampini 4715e7931f94SStefano Zampini /* Get info on mapping */ 47163bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 47173bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4718e7931f94SStefano Zampini 4719e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4720785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4721e7931f94SStefano Zampini xadj[0] = 0; 4722e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4723785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4724785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4725e7931f94SStefano Zampini 47262b510759SStefano Zampini if (threshold) { 4727d023bfaeSStefano Zampini PetscInt xadj_count = 0; 47282b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4729d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4730d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4731d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4732d023bfaeSStefano Zampini xadj_count++; 4733e7931f94SStefano Zampini } 4734e7931f94SStefano Zampini } 4735d023bfaeSStefano Zampini xadj[1] = xadj_count; 4736c8587f34SStefano Zampini } else { 4737e7931f94SStefano Zampini if (xadj[1]) { 4738e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4739e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4740c8587f34SStefano Zampini } 4741e7931f94SStefano Zampini } 47423bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4743e7931f94SStefano Zampini if (use_square) { 4744e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4745e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4746e7931f94SStefano Zampini } 4747e7931f94SStefano Zampini } 4748e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4749e7931f94SStefano Zampini 47503837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4751e7931f94SStefano Zampini 4752e7931f94SStefano Zampini /* 4753e7931f94SStefano Zampini Restrict work on active processes only. 4754e7931f94SStefano Zampini */ 4755e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4756e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4757e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 47582b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4759d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4760e7931f94SStefano Zampini if (color) { 4761e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4762e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4763e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4764c8587f34SStefano Zampini } else { 476552e5ac9dSStefano Zampini Mat subdomain_adj; 476652e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 476752e5ac9dSStefano Zampini MatPartitioning partitioner; 476852e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 476952e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4770b0c7d250SStefano Zampini PetscBool aggregate; 4771b0c7d250SStefano Zampini 4772306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4773785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4774e7931f94SStefano Zampini prank = rank; 4775306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 47768002ef2cSStefano Zampini /* 4777e7931f94SStefano Zampini for (i=0;i<size;i++) { 4778e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4779c8587f34SStefano Zampini } 47808002ef2cSStefano Zampini */ 4781e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4782e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4783c8587f34SStefano Zampini } 4784e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4785b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4786b0c7d250SStefano Zampini if (aggregate) { 4787b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4788b0c7d250SStefano Zampini PetscMPIInt nrank; 4789b0c7d250SStefano Zampini PetscScalar *vals; 4790b0c7d250SStefano Zampini 4791b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4792b0c7d250SStefano Zampini lrows = 0; 4793b0c7d250SStefano Zampini if (nrank<redprocs) { 4794b0c7d250SStefano Zampini lrows = size/redprocs; 4795b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4796b0c7d250SStefano Zampini } 47975fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4798b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4799b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4800b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4801b0c7d250SStefano Zampini row = nrank; 4802b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4803b0c7d250SStefano Zampini cols = adjncy; 4804b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4805b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4806b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4807b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4808b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 480952e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 481052e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 481152e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4812b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4813b0c7d250SStefano Zampini } else { 4814306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4815b0c7d250SStefano Zampini } 481622b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4817e7931f94SStefano Zampini 4818e7931f94SStefano Zampini /* Partition */ 4819306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4820e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4821e7931f94SStefano Zampini if (use_vwgt) { 48223837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4823e7931f94SStefano Zampini v_wgt[0] = local_size; 4824e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4825c8587f34SStefano Zampini } 482628143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 482728143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4828e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4829e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 483022b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4831e7931f94SStefano Zampini 483252e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 483352e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 483452e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 483552e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4836b0c7d250SStefano Zampini if (!redprocs) { 4837b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 483828143c3dSStefano Zampini } else { 4839b0c7d250SStefano Zampini PetscInt idxs[1]; 4840b0c7d250SStefano Zampini PetscMPIInt tag; 4841b0c7d250SStefano Zampini MPI_Request *reqs; 4842b0c7d250SStefano Zampini 4843b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4844b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4845b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4846b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 484728143c3dSStefano Zampini } 4848b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4849b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4850b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4851b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4852e7931f94SStefano Zampini } 485352e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4854e7931f94SStefano Zampini /* clean up */ 4855e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 485652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4857e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4858e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4859e7931f94SStefano Zampini } 4860e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4861e7931f94SStefano Zampini 4862e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4863e7931f94SStefano Zampini i = 1; 4864e7931f94SStefano Zampini if (color) i=0; 4865e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4866e7931f94SStefano Zampini /* get back IS */ 4867e7931f94SStefano Zampini *is_sends = ranks_send_to; 4868e7931f94SStefano Zampini PetscFunctionReturn(0); 4869e7931f94SStefano Zampini } 4870e7931f94SStefano Zampini 4871e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4872e7931f94SStefano Zampini 4873e7931f94SStefano Zampini #undef __FUNCT__ 4874e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 487553a05cb3SStefano 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[]) 4876e7931f94SStefano Zampini { 487770cf5478SStefano Zampini Mat local_mat; 4878e7931f94SStefano Zampini IS is_sends_internal; 48799d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 488028143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 48819d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4882e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4883e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4884e7931f94SStefano Zampini const PetscInt* is_indices; 4885e7931f94SStefano Zampini MatType new_local_type; 4886e7931f94SStefano Zampini /* buffers */ 4887e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 488828143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 48899d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4890e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4891e7931f94SStefano Zampini /* MPI */ 489228143c3dSStefano Zampini MPI_Comm comm,comm_n; 489328143c3dSStefano Zampini PetscSubcomm subcomm; 4894e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 489528143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 489628143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 489728143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 489828143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 489928143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4900e7931f94SStefano Zampini PetscErrorCode ierr; 4901e7931f94SStefano Zampini 4902e7931f94SStefano Zampini PetscFunctionBegin; 490328143c3dSStefano Zampini /* TODO: add missing checks */ 490428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 490528143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 490628143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 490728143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4908e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 490928143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4910e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4911e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4912e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4913e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4914e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 491528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 491670cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 491770cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 491828143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 491970cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 492070cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 492170cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 492270cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 492370cf5478SStefano Zampini } 4924e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4925e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4926e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4927e7931f94SStefano Zampini if (!is_sends) { 492828143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4929b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4930c8587f34SStefano Zampini } else { 4931e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4932e7931f94SStefano Zampini is_sends_internal = is_sends; 4933c8587f34SStefano Zampini } 4934e7931f94SStefano Zampini 4935e7931f94SStefano Zampini /* get comm */ 4936a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4937e7931f94SStefano Zampini 4938e7931f94SStefano Zampini /* compute number of sends */ 4939e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4940e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4941e7931f94SStefano Zampini 4942e7931f94SStefano Zampini /* compute number of receives */ 4943e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4944785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4945e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4946e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4947e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4948e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4949e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4950e7931f94SStefano Zampini 495128143c3dSStefano Zampini /* restrict comm if requested */ 495228143c3dSStefano Zampini subcomm = 0; 495328143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 495428143c3dSStefano Zampini if (restrict_comm) { 4955779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4956779c1cceSStefano Zampini 495728143c3dSStefano Zampini color = 0; 495853a05cb3SStefano Zampini if (restrict_full) { 495953a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 496053a05cb3SStefano Zampini } else { 496153a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 496253a05cb3SStefano Zampini } 4963b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 496428143c3dSStefano Zampini subcommsize = commsize - subcommsize; 496528143c3dSStefano Zampini /* check if reuse has been requested */ 496628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 496728143c3dSStefano Zampini if (*mat_n) { 496828143c3dSStefano Zampini PetscMPIInt subcommsize2; 496928143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 497028143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 497128143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 497228143c3dSStefano Zampini } else { 497328143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 497428143c3dSStefano Zampini } 497528143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4976779c1cceSStefano Zampini PetscMPIInt rank; 4977779c1cceSStefano Zampini 4978779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 497928143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 498028143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 498128143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4982306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 498328143c3dSStefano Zampini } 498428143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 498528143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 498628143c3dSStefano Zampini } else { 498728143c3dSStefano Zampini comm_n = comm; 498828143c3dSStefano Zampini } 498928143c3dSStefano Zampini 4990e7931f94SStefano Zampini /* prepare send/receive buffers */ 4991785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4992e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4993785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4994e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 499528143c3dSStefano Zampini if (nis) { 4996854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 499728143c3dSStefano Zampini } 4998e7931f94SStefano Zampini 499928143c3dSStefano Zampini /* Get data from local matrices */ 50006c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5001e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5002e7931f94SStefano Zampini /* 5003e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5004e7931f94SStefano Zampini send_buffer_idxs should contain: 5005e7931f94SStefano Zampini - MatType_PRIVATE type 5006e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5007e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5008e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5009e7931f94SStefano Zampini */ 50106c4ed002SBarry Smith else { 5011e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 50123bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5013854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5014e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5015e7931f94SStefano Zampini send_buffer_idxs[1] = i; 50163bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5017e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 50183bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5019e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5020e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5021e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5022e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5023c8587f34SStefano Zampini } 5024c8587f34SStefano Zampini } 5025e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 502628143c3dSStefano Zampini /* additional is (if any) */ 502728143c3dSStefano Zampini if (nis) { 502828143c3dSStefano Zampini PetscMPIInt psum; 502928143c3dSStefano Zampini PetscInt j; 503028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 503128143c3dSStefano Zampini PetscInt plen; 503228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 503328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 503428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 503528143c3dSStefano Zampini } 5036854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 503728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 503828143c3dSStefano Zampini PetscInt plen; 503928143c3dSStefano Zampini const PetscInt *is_array_idxs; 504028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 504128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 504228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 504328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 504428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 504528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 504628143c3dSStefano Zampini } 504728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 504828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 504928143c3dSStefano Zampini } 505028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 505128143c3dSStefano Zampini } 505228143c3dSStefano Zampini 5053e7931f94SStefano Zampini buf_size_idxs = 0; 5054e7931f94SStefano Zampini buf_size_vals = 0; 505528143c3dSStefano Zampini buf_size_idxs_is = 0; 5056e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5057e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5058e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 505928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5060e7931f94SStefano Zampini } 5061785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5062785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 506395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5064e7931f94SStefano Zampini 5065e7931f94SStefano Zampini /* get new tags for clean communications */ 5066e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5067e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 506828143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5069e7931f94SStefano Zampini 5070e7931f94SStefano Zampini /* allocate for requests */ 5071785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5072785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 507395ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5074785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5075785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 507695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5077e7931f94SStefano Zampini 5078e7931f94SStefano Zampini /* communications */ 5079e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5080e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 508128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5082e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5083e7931f94SStefano Zampini source_dest = onodes[i]; 5084e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5085e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5086e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5087e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 508828143c3dSStefano Zampini if (nis) { 508928143c3dSStefano 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); 509028143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 509128143c3dSStefano Zampini } 5092e7931f94SStefano Zampini } 5093e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5094e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5095e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5096e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 509728143c3dSStefano Zampini if (nis) { 509828143c3dSStefano 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); 509928143c3dSStefano Zampini } 5100e7931f94SStefano Zampini } 5101e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5102e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5103e7931f94SStefano Zampini 5104e7931f94SStefano Zampini /* assemble new l2g map */ 5105e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5106e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51079d30be91SStefano Zampini new_local_rows = 0; 5108e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51099d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5110e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5111e7931f94SStefano Zampini } 51129d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5113e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51149d30be91SStefano Zampini new_local_rows = 0; 5115e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51169d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 51179d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5118e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5119e7931f94SStefano Zampini } 51209d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 51219d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5122e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5123e7931f94SStefano Zampini 5124e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5125e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5126e7931f94SStefano 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) */ 5127e7931f94SStefano Zampini if (n_recvs) { 512828143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5129e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5130e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5131e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5132e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5133e7931f94SStefano Zampini break; 5134e7931f94SStefano Zampini } 5135e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5136e7931f94SStefano Zampini } 5137e7931f94SStefano Zampini switch (new_local_type_private) { 513828143c3dSStefano Zampini case MATDENSE_PRIVATE: 513928143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5140e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5141e7931f94SStefano Zampini bs = 1; 514228143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 514328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 514428143c3dSStefano Zampini bs = 1; 514528143c3dSStefano Zampini } 5146e7931f94SStefano Zampini break; 5147e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5148e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5149e7931f94SStefano Zampini bs = 1; 5150e7931f94SStefano Zampini break; 5151e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5152e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5153e7931f94SStefano Zampini break; 5154e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5155e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5156e7931f94SStefano Zampini break; 5157e7931f94SStefano Zampini default: 51589d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5159e7931f94SStefano Zampini break; 5160e7931f94SStefano Zampini } 516128143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 516228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 516328143c3dSStefano Zampini bs = 1; 5164e7931f94SStefano Zampini } 5165e7931f94SStefano Zampini 516670cf5478SStefano Zampini /* create MATIS object if needed */ 516770cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5168e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5169e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 517070cf5478SStefano Zampini } else { 517170cf5478SStefano Zampini /* it also destroys the local matrices */ 517270cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 517370cf5478SStefano Zampini } 517470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5175e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 51769d30be91SStefano Zampini 51779d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 51789d30be91SStefano Zampini 51799d30be91SStefano Zampini /* Global to local map of received indices */ 51809d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 51819d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 51829d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 51839d30be91SStefano Zampini 51849d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 51859d30be91SStefano Zampini buf_size_idxs = 0; 51869d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51879d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 51889d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 51899d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 51909d30be91SStefano Zampini } 51919d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 51929d30be91SStefano Zampini 51939d30be91SStefano Zampini /* set preallocation */ 51949d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 51959d30be91SStefano Zampini if (!newisdense) { 51969d30be91SStefano Zampini PetscInt *new_local_nnz=0; 51979d30be91SStefano Zampini 51989d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 51999d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 52009d30be91SStefano Zampini if (n_recvs) { 52019d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 52029d30be91SStefano Zampini } 52039d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 52049d30be91SStefano Zampini PetscInt j; 52059d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 52069d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 52079d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 52089d30be91SStefano Zampini } 52099d30be91SStefano Zampini } else { 52109d30be91SStefano Zampini /* TODO */ 52119d30be91SStefano Zampini } 52129d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 52139d30be91SStefano Zampini } 52149d30be91SStefano Zampini if (new_local_nnz) { 52159d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 52169d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 52179d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 52189d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52199d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 52209d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52219d30be91SStefano Zampini } else { 52229d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52239d30be91SStefano Zampini } 52249d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 52259d30be91SStefano Zampini } else { 52269d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52279d30be91SStefano Zampini } 5228e7931f94SStefano Zampini 5229e7931f94SStefano Zampini /* set values */ 5230e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 52319d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5232e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5233e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5234e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 52359d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5236e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5237e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5238e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 523928143c3dSStefano Zampini } else { 524028143c3dSStefano Zampini /* TODO */ 5241e7931f94SStefano Zampini } 5242e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5243e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5244e7931f94SStefano Zampini } 5245e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5246e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 524770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 524870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 52499d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 52509d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5251e7931f94SStefano Zampini 5252dfd14d43SStefano Zampini #if 0 525328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5254e7931f94SStefano Zampini Vec lvec,rvec; 5255e7931f94SStefano Zampini PetscReal infty_error; 5256e7931f94SStefano Zampini 52572a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5258e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5259e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5260e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 526170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5262e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5263e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5264e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5265e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5266e7931f94SStefano Zampini } 526728143c3dSStefano Zampini #endif 5268e7931f94SStefano Zampini 526928143c3dSStefano Zampini /* assemble new additional is (if any) */ 527028143c3dSStefano Zampini if (nis) { 527128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 527228143c3dSStefano Zampini 527328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5274854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 527528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 527628143c3dSStefano Zampini psum = 0; 527728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 527828143c3dSStefano Zampini for (j=0;j<nis;j++) { 527928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 528028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 528128143c3dSStefano Zampini psum += plen; 528228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 528328143c3dSStefano Zampini } 528428143c3dSStefano Zampini } 5285854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5286854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 528728143c3dSStefano Zampini for (i=1;i<nis;i++) { 528828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 528928143c3dSStefano Zampini } 529028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 529128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 529228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 529328143c3dSStefano Zampini for (j=0;j<nis;j++) { 529428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 529528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 529628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 529728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 529828143c3dSStefano Zampini } 529928143c3dSStefano Zampini } 530028143c3dSStefano Zampini for (i=0;i<nis;i++) { 530128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 530228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 530328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 530428143c3dSStefano Zampini } 530528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 530628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 530728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 530828143c3dSStefano Zampini } 5309e7931f94SStefano Zampini /* free workspace */ 531028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5311e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5312e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5313e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5314e7931f94SStefano Zampini if (isdense) { 5315e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5316e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5317e7931f94SStefano Zampini } else { 5318e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5319e7931f94SStefano Zampini } 532028143c3dSStefano Zampini if (nis) { 532128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 532228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 532328143c3dSStefano Zampini } 5324e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5325e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 532628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5327e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5328e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 532928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5330e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5331e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5332e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5333e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5334e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 533528143c3dSStefano Zampini if (nis) { 533628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 533728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 533828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 533928143c3dSStefano Zampini } 534028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 534128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 534228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 534328143c3dSStefano Zampini for (i=0;i<nis;i++) { 534428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 534528143c3dSStefano Zampini } 534653a05cb3SStefano Zampini *mat_n = NULL; 534728143c3dSStefano Zampini } 5348e7931f94SStefano Zampini PetscFunctionReturn(0); 5349e7931f94SStefano Zampini } 5350a57a6d2fSStefano Zampini 535112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5352af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 535312edc857SStefano Zampini 5354c8587f34SStefano Zampini #undef __FUNCT__ 5355c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5356c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5357c8587f34SStefano Zampini { 5358c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5359c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 536020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 53619881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 536220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 53636e683305SStefano Zampini IS coarse_is,*isarray; 53646e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 536530368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5366f9eb5b7dSStefano Zampini PC pc_temp; 5367c8587f34SStefano Zampini PCType coarse_pc_type; 5368c8587f34SStefano Zampini KSPType coarse_ksp_type; 5369f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 53704f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 53716e683305SStefano Zampini Mat t_coarse_mat_is; 53726e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 53736e683305SStefano Zampini PetscMPIInt all_procs; 537474e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 537568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 537622bc73bbSStefano Zampini PetscScalar *array; 53779881197aSStefano Zampini PetscErrorCode ierr; 5378fdc09c96SStefano Zampini 5379c8587f34SStefano Zampini PetscFunctionBegin; 5380c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 538168457ee5SStefano 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 */ 5382fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 53835a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5384fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5385f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5386f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5387f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5388fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 538951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 539051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5391dc4bcba2SStefano Zampini PC pc; 5392dc4bcba2SStefano Zampini PetscBool isbddc; 5393dc4bcba2SStefano Zampini 5394dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5395dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5396dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5397dc4bcba2SStefano Zampini if (isbddc) { 539863c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 539963c961adSStefano Zampini } else { 5400727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 540163c961adSStefano Zampini } 5402fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5403fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5404fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5405f4ddd8eeSStefano Zampini } 5406fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5407fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5408f4ddd8eeSStefano Zampini } 540970cf5478SStefano Zampini /* reset any subassembling information */ 541070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 54116e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54126e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5413fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5414f4ddd8eeSStefano Zampini } 5415c8587f34SStefano Zampini 54166e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 54172b510759SStefano Zampini im_active = !!(pcis->n); 5418b2566f29SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 54196e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 54206e683305SStefano Zampini void_procs = all_procs-active_procs; 54216e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 542274e2c79eSStefano Zampini redist = PETSC_FALSE; 542322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 54246e683305SStefano Zampini csin_ml = PETSC_TRUE; 54256e683305SStefano Zampini ncoarse_ml = void_procs; 5426779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5427779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 54286e683305SStefano Zampini csin_ds = PETSC_TRUE; 542918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 543018a45a71SStefano Zampini redist = PETSC_TRUE; 543118a45a71SStefano Zampini } else { 54326e683305SStefano Zampini csin_ds = PETSC_TRUE; 5433779c1cceSStefano Zampini ncoarse_ds = active_procs; 5434779c1cceSStefano Zampini redist = PETSC_TRUE; 543518a45a71SStefano Zampini } 54366e683305SStefano Zampini } else { 54376e683305SStefano Zampini csin_ml = PETSC_FALSE; 54386e683305SStefano Zampini ncoarse_ml = all_procs; 54396e683305SStefano Zampini if (void_procs) { 54406e683305SStefano Zampini csin_ds = PETSC_TRUE; 54416e683305SStefano Zampini ncoarse_ds = void_procs; 54426e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 54436e683305SStefano Zampini } else { 5444779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 544574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 544674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 544774e2c79eSStefano Zampini redist = PETSC_TRUE; 544874e2c79eSStefano Zampini } else { 54496e683305SStefano Zampini csin_ds = PETSC_FALSE; 54506e683305SStefano Zampini ncoarse_ds = all_procs; 54516e683305SStefano Zampini } 54526e683305SStefano Zampini } 545374e2c79eSStefano Zampini } 54546e683305SStefano Zampini 54556e683305SStefano Zampini /* 54566e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 54576e683305SStefano Zampini - we have not exceeded the number of levels requested 54586e683305SStefano Zampini - we can actually subassemble the active processes 54596e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 54606e683305SStefano Zampini */ 54616e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 54626e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 54636e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 54646e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 54656e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5466f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 54672b510759SStefano Zampini } else { 5468f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5469c8587f34SStefano Zampini } 5470c8587f34SStefano Zampini } 54716e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 54726e683305SStefano Zampini if (multilevel_allowed) { 54736e683305SStefano Zampini ncoarse = ncoarse_ml; 54746e683305SStefano Zampini csin = csin_ml; 547558da7f69SStefano Zampini redist = PETSC_FALSE; 54766e683305SStefano Zampini } else { 54776e683305SStefano Zampini ncoarse = ncoarse_ds; 54786e683305SStefano Zampini csin = csin_ds; 54796e683305SStefano Zampini } 5480e7931f94SStefano Zampini 5481abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5482abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5483abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5484abbbba34SStefano Zampini 5485abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 548622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 548722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 548822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 548922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5490e176bc59SStefano 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); 54916e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 54926e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 54936e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5494abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5495abbbba34SStefano Zampini 54966e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 549730368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 54986e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 54996e683305SStefano Zampini const PetscInt *idxs; 55006e683305SStefano Zampini ISLocalToGlobalMapping tmap; 55016e683305SStefano Zampini 55026e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 55030be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 55046e683305SStefano Zampini /* allocate space for temporary storage */ 5505854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5506854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 55076e683305SStefano Zampini /* allocate for IS array */ 55086e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 55096e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 551030368db7SStefano Zampini nisvert = 0; 551130368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 551230368db7SStefano Zampini nisvert = 1; 551330368db7SStefano Zampini } 551430368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5515854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 55166e683305SStefano Zampini /* dofs splitting */ 55176e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 55186e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 55196e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 55206e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55216e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55226e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55236e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 552430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 55256e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 55266e683305SStefano Zampini } 55276e683305SStefano Zampini /* neumann boundaries */ 55286e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 55296e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 55306e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 55316e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55326e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55336e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55346e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 553530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 55366e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 55376e683305SStefano Zampini } 553830368db7SStefano Zampini /* primal vertices (benign) */ 553930368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 554030368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 554130368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 554230368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 554330368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 554430368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 554530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 554630368db7SStefano Zampini } 55476e683305SStefano Zampini /* free memory */ 55486e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 55496e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 55506e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 55516e683305SStefano Zampini } else { 55526e683305SStefano Zampini nis = 0; 55536e683305SStefano Zampini nisdofs = 0; 55546e683305SStefano Zampini nisneu = 0; 555530368db7SStefano Zampini nisvert = 0; 55566e683305SStefano Zampini isarray = NULL; 55576e683305SStefano Zampini } 55586e683305SStefano Zampini /* destroy no longer needed map */ 55596e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 55606e683305SStefano Zampini 55616e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 55626e683305SStefano Zampini coarse_mat_is = NULL; 55636e683305SStefano Zampini if (csin) { 55646e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 556574e2c79eSStefano Zampini if (redist) { 556674e2c79eSStefano Zampini PetscMPIInt rank; 5567779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 556874e2c79eSStefano Zampini 556974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 557058da7f69SStefano Zampini spc = active_procs/ncoarse; 557158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5572779c1cceSStefano Zampini if (im_active) { 5573779c1cceSStefano Zampini destsize = 1; 557474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 557574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 557674e2c79eSStefano Zampini } else { 557774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 557874e2c79eSStefano Zampini } 557974e2c79eSStefano Zampini } else { 5580779c1cceSStefano Zampini destsize = 0; 55816e683305SStefano Zampini } 5582779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5583779c1cceSStefano Zampini } else if (csin_type_simple) { 55846e683305SStefano Zampini PetscMPIInt rank; 55856e683305SStefano Zampini PetscInt issize,isidx; 5586779c1cceSStefano Zampini 55876e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 55886e683305SStefano Zampini if (im_active) { 55896e683305SStefano Zampini issize = 1; 55906e683305SStefano Zampini isidx = (PetscInt)rank; 55916e683305SStefano Zampini } else { 55926e683305SStefano Zampini issize = 0; 55936e683305SStefano Zampini isidx = -1; 55946e683305SStefano Zampini } 55956e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5596779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5597b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 55986e683305SStefano Zampini } 5599779c1cceSStefano Zampini 5600779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5601779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5602779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5603779c1cceSStefano Zampini PetscInt *coarse_candidates; 5604779c1cceSStefano Zampini const PetscInt* tisindices; 5605779c1cceSStefano Zampini 5606779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5607779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5608779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5609779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5610779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5611779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5612779c1cceSStefano Zampini } 5613779c1cceSStefano Zampini } 5614779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5615779c1cceSStefano Zampini 5616779c1cceSStefano Zampini 56176e683305SStefano Zampini if (pcbddc->dbg_flag) { 56186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 56206e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5622779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 56236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 56246e683305SStefano Zampini } 56256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 56266e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56276e683305SStefano Zampini } 56286e683305SStefano Zampini /* shift the pattern on coarse candidates */ 56296e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 56306e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5631854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 56326e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 56336e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 56346e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 56356e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 56366e683305SStefano Zampini } 56376e683305SStefano Zampini if (pcbddc->dbg_flag) { 56386e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 56406e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56416e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56426e683305SStefano Zampini } 5643779c1cceSStefano Zampini } 56446e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 564553a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 564653a05cb3SStefano 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); 564753a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 564853a05cb3SStefano 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); 564953a05cb3SStefano Zampini } 56506e683305SStefano Zampini } else { 56516e683305SStefano Zampini if (pcbddc->dbg_flag) { 56526e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 56546e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56556e683305SStefano Zampini } 56566e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 56576e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 56586e683305SStefano Zampini } 56596e683305SStefano Zampini 56606e683305SStefano Zampini /* create local to global scatters for coarse problem */ 566168457ee5SStefano Zampini if (compute_vecs) { 56626e683305SStefano Zampini PetscInt lrows; 56636e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 56646e683305SStefano Zampini if (coarse_mat_is) { 56656e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 56666e683305SStefano Zampini } else { 56676e683305SStefano Zampini lrows = 0; 56686e683305SStefano Zampini } 56696e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 56706e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 56716e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 56726e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56736e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56746e683305SStefano Zampini } 56756e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 56766e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5677c8587f34SStefano Zampini 5678f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5679f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5680f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5681f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5682f9eb5b7dSStefano Zampini } else { 5683f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5684f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5685c8587f34SStefano Zampini } 5686c8587f34SStefano Zampini 56876e683305SStefano Zampini /* print some info if requested */ 56886e683305SStefano Zampini if (pcbddc->dbg_flag) { 56896e683305SStefano Zampini if (!multilevel_allowed) { 56906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56916e683305SStefano Zampini if (multilevel_requested) { 56926e683305SStefano 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); 56936e683305SStefano Zampini } else if (pcbddc->max_levels) { 56946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 56956e683305SStefano Zampini } 56966e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56976e683305SStefano Zampini } 56986e683305SStefano Zampini } 56996e683305SStefano Zampini 5700f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 57016e683305SStefano Zampini if (coarse_mat_is) { 57026e683305SStefano Zampini MatReuse coarse_mat_reuse; 57036a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 57046e683305SStefano Zampini if (pcbddc->dbg_flag) { 57056e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 57066e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57076e683305SStefano Zampini } 5708f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5709312be037SStefano Zampini char prefix[256],str_level[16]; 5710e604994aSStefano Zampini size_t len; 57116e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5712422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5713c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5714f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 57155f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5716c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 57176e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5718c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5719c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5720e604994aSStefano Zampini /* prefix */ 5721e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5722e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5723e604994aSStefano Zampini if (!pcbddc->current_level) { 5724e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5725e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5726c8587f34SStefano Zampini } else { 5727e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5728312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5729312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 573034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5731312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5732e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5733e604994aSStefano Zampini } 5734e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 57353e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 57363e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 57373e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 57383e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5739f9eb5b7dSStefano Zampini /* allow user customization */ 5740f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 57413e3c6dadSStefano Zampini } 57423e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 574351bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 57443e3c6dadSStefano Zampini if (nisdofs) { 57453e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 57463e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 57473e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 57483e3c6dadSStefano Zampini } 57493e3c6dadSStefano Zampini } 57503e3c6dadSStefano Zampini if (nisneu) { 57513e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 57523e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5753312be037SStefano Zampini } 575430368db7SStefano Zampini if (nisvert) { 575530368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 575630368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 575730368db7SStefano Zampini } 5758f9eb5b7dSStefano Zampini 5759f9eb5b7dSStefano Zampini /* get some info after set from options */ 5760f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5761f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 57624f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 57636e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5764f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5765f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5766f9eb5b7dSStefano Zampini } 576739f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 57684f3a063dSStefano Zampini if (isredundant) { 57694f3a063dSStefano Zampini KSP inner_ksp; 57704f3a063dSStefano Zampini PC inner_pc; 57714f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 57724f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 57734f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 57744f3a063dSStefano Zampini } 5775f9eb5b7dSStefano Zampini 5776f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5777fa7f1dd8SStefano Zampini if (coarse_reuse) { 577881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5779fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 57806e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5781fa7f1dd8SStefano Zampini } else { 57826e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5783fa7f1dd8SStefano Zampini } 5784c8587f34SStefano Zampini if (isbddc || isnn) { 5785720d30f9SStefano Zampini if (isbddc) { /* currently there are no APIs for these options */ 5786720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5787720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 5788720d30f9SStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_saddle_point; 5789720d30f9SStefano Zampini pcbddc_coarse->benign_compute_nonetflux = pcbddc->benign_compute_nonetflux; 5790720d30f9SStefano Zampini if (pcbddc_coarse->benign_compute_nonetflux) { 5791720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 5792720d30f9SStefano Zampini } 5793d4d8cf7bSStefano Zampini } 579422bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 579570cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5796b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 579722b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 57986e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 58006e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 58016e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 580222b6e8a2SStefano Zampini } 580370cf5478SStefano Zampini } 580453a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 580570cf5478SStefano Zampini } else { 580622bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 580722bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 580822bc73bbSStefano Zampini } 580922bc73bbSStefano Zampini } else { 58102e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5811c8587f34SStefano Zampini } 5812c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5813c8587f34SStefano Zampini 58143301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 58155a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 58163301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 58173301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 58183301b35fSStefano Zampini } 58193301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 58203301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 58213301b35fSStefano Zampini } 58223301b35fSStefano Zampini if (pc->pmat->spd_set) { 58233301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 58243301b35fSStefano Zampini } 58256e683305SStefano Zampini /* set operators */ 58265f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 58276e683305SStefano Zampini if (pcbddc->dbg_flag) { 58286e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 58296e683305SStefano Zampini } 58306e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 58316e683305SStefano Zampini coarse_mat = 0; 58326e683305SStefano Zampini } 58336e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5834b1ecc7b1SStefano Zampini #if 0 5835b9b85e73SStefano Zampini { 5836b9b85e73SStefano Zampini PetscViewer viewer; 5837b9b85e73SStefano Zampini char filename[256]; 5838b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5839b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 58406a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5841b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5842f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 5843b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5844b9b85e73SStefano Zampini } 5845b9b85e73SStefano Zampini #endif 5846c8587f34SStefano Zampini 5847c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5848298c0119SStefano Zampini #if 0 5849c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5850c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 585198a51de6SStefano Zampini } 5852298c0119SStefano Zampini #endif 5853b0f5fe93SStefano Zampini /* hack */ 585498a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 585598a51de6SStefano Zampini Vec crhs,csol; 585604708bb6SStefano Zampini 5857f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5858f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5859f347579bSStefano Zampini if (!csol) { 58602a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5861f9eb5b7dSStefano Zampini } 5862f347579bSStefano Zampini if (!crhs) { 58632a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5864f347579bSStefano Zampini } 5865b0f5fe93SStefano Zampini } 5866b0f5fe93SStefano Zampini 5867b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5868b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5869b0f5fe93SStefano Zampini 5870b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 58714f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58724f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 58734f1b2e48SStefano Zampini } 5874b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5875b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5876b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5877b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5878b0f5fe93SStefano Zampini if (coarse_mat) { 5879b0f5fe93SStefano Zampini Vec nullv; 5880b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5881b0f5fe93SStefano Zampini PetscInt nl; 5882b0f5fe93SStefano Zampini 5883b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5884b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5885b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5886b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5887b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5888b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5889b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5890b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5891b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5892b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5893b0f5fe93SStefano Zampini } 5894b0f5fe93SStefano Zampini } 5895b0f5fe93SStefano Zampini 5896b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5897b0f5fe93SStefano Zampini PetscBool ispreonly; 5898b0f5fe93SStefano Zampini 5899b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5900b0f5fe93SStefano Zampini PetscBool isnull; 5901b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5902d4d8cf7bSStefano Zampini if (0) { 590330368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5904b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5905b0f5fe93SStefano Zampini } else { 5906b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5907b0f5fe93SStefano Zampini } 5908b0f5fe93SStefano Zampini } else { 5909b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5910b0f5fe93SStefano Zampini } 5911b0f5fe93SStefano Zampini } 5912b0f5fe93SStefano Zampini /* setup coarse ksp */ 5913b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5914cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5915cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 59166e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5917c8587f34SStefano Zampini KSP check_ksp; 59182b510759SStefano Zampini KSPType check_ksp_type; 5919c8587f34SStefano Zampini PC check_pc; 59206e683305SStefano Zampini Vec check_vec,coarse_vec; 59216a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 59222b510759SStefano Zampini PetscInt its; 59236e683305SStefano Zampini PetscBool compute_eigs; 59246e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 59256e683305SStefano Zampini PetscInt neigs; 59268e185a42SStefano Zampini const char *prefix; 5927c8587f34SStefano Zampini 59282b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 59296e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5930422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 593123ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5932f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 59332b510759SStefano Zampini if (ispreonly) { 59342b510759SStefano Zampini check_ksp_type = KSPPREONLY; 59356e683305SStefano Zampini compute_eigs = PETSC_FALSE; 59362b510759SStefano Zampini } else { 5937cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 59386e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5939c8587f34SStefano Zampini } 5940c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 59416e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 59426e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 59436e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5944a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5945a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5946a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5947a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5948c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5949c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5950c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5951c8587f34SStefano Zampini /* create random vec */ 59526e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 59536e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5954c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5955c8587f34SStefano Zampini if (CoarseNullSpace) { 5956c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5957c8587f34SStefano Zampini } 59586e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5959c8587f34SStefano Zampini /* solve coarse problem */ 59606e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5961c8587f34SStefano Zampini if (CoarseNullSpace) { 59626e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5963c8587f34SStefano Zampini } 5964cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 59656e683305SStefano Zampini if (compute_eigs) { 5966854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5967854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 59686e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 59696e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 59706e683305SStefano Zampini lambda_min = eigs_r[0]; 59716e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 59726e683305SStefano Zampini if (lambda_max>lambda_min) { 5973cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5974cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5975cbcc2c2aSStefano Zampini } 5976c8587f34SStefano Zampini } 5977c8587f34SStefano Zampini } 5978cbcc2c2aSStefano Zampini 5979c8587f34SStefano Zampini /* check coarse problem residual error */ 59806e683305SStefano Zampini if (pcbddc->dbg_flag) { 59816e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 59826e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59836e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5984c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 59856e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 59866e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5987c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5988779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 59896e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 59906e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 59916e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 59926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5993b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5994b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5995b0f5fe93SStefano Zampini } 59966e683305SStefano Zampini if (compute_eigs) { 59976e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5998deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5999c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 60006e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 60016e683305SStefano 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); 60026e683305SStefano Zampini for (i=0;i<neigs;i++) { 60036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6004c8587f34SStefano Zampini } 60056e683305SStefano Zampini } 60066e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 60076e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 60086e683305SStefano Zampini } 6009c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 60106e683305SStefano Zampini if (compute_eigs) { 60116e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 60126e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6013c8587f34SStefano Zampini } 60146e683305SStefano Zampini } 60156e683305SStefano Zampini } 6016cbcc2c2aSStefano Zampini /* print additional info */ 6017cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 60186e683305SStefano Zampini /* waits until all processes reaches this point */ 60196e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6020cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6021cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6022cbcc2c2aSStefano Zampini } 6023cbcc2c2aSStefano Zampini 60242b510759SStefano Zampini /* free memory */ 6025c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6026fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6027c8587f34SStefano Zampini PetscFunctionReturn(0); 6028c8587f34SStefano Zampini } 6029674ae819SStefano Zampini 6030f34684f1SStefano Zampini #undef __FUNCT__ 6031f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6032f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6033f34684f1SStefano Zampini { 6034f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6035f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6036f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6037dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6038dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 603973be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6040dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6041f34684f1SStefano Zampini PetscErrorCode ierr; 6042f34684f1SStefano Zampini 6043f34684f1SStefano Zampini PetscFunctionBegin; 6044f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 60456c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6046dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 60473bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6048dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6049dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6050dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6051dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6052dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6053dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60546c4ed002SBarry 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); 6055dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6056dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6057dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6058dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6059dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6060f34684f1SStefano Zampini 6061f34684f1SStefano Zampini /* check numbering */ 6062f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6063019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6064dc456d91SStefano Zampini PetscInt i; 6065b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6066f34684f1SStefano Zampini 6067f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6068f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6069f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 60701575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6071019a44ceSStefano Zampini /* counter */ 6072019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6073019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6074019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6075019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6076019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6077019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6078f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6079f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6080727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6081f34684f1SStefano Zampini } 6082f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6083f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6084f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6085e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6086e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6087e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6088e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6089f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6090019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6091f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6092019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 609375c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 609475c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6095b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6096019a44ceSStefano 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); 6097f34684f1SStefano Zampini } 6098f34684f1SStefano Zampini } 6099019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6100b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6101f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6102f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6103f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6104f34684f1SStefano Zampini } 6105f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6106f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6107e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6108e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6109f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6110f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6111b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6112ca8b9ea9SStefano Zampini PetscInt *gidxs; 6113ca8b9ea9SStefano Zampini 6114ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 61153bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6116f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6117f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6118f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6119f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 61204bc2dc4bSStefano 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); 6121f34684f1SStefano Zampini } 6122f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6123ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6124f34684f1SStefano Zampini } 6125f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 61261575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6127302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6128f34684f1SStefano Zampini } 61298bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6130f34684f1SStefano Zampini /* get back data */ 6131f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6132f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6133674ae819SStefano Zampini PetscFunctionReturn(0); 6134674ae819SStefano Zampini } 6135674ae819SStefano Zampini 6136e456f2a8SStefano Zampini #undef __FUNCT__ 6137e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6138a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6139e456f2a8SStefano Zampini { 6140e456f2a8SStefano Zampini IS localis_t; 6141a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6142e456f2a8SStefano Zampini PetscScalar *vals; 6143e456f2a8SStefano Zampini PetscErrorCode ierr; 6144e456f2a8SStefano Zampini 6145e456f2a8SStefano Zampini PetscFunctionBegin; 6146a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6147e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6148854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6149e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6150e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6151a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6152a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61531035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6154a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61551035eff8SStefano Zampini } 6156a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6157e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6158e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6159a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6160a7dc3881SStefano Zampini /* now compute set in local ordering */ 6161a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6162a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6163a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6164a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6165a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6166ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6167e456f2a8SStefano Zampini lsize++; 6168e456f2a8SStefano Zampini } 6169e456f2a8SStefano Zampini } 6170854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6171a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6172ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6173e456f2a8SStefano Zampini idxs[lsize++] = i; 6174e456f2a8SStefano Zampini } 6175e456f2a8SStefano Zampini } 6176a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6177a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6178e456f2a8SStefano Zampini *localis = localis_t; 6179e456f2a8SStefano Zampini PetscFunctionReturn(0); 6180e456f2a8SStefano Zampini } 6181906d46d4SStefano Zampini 6182b96c3477SStefano Zampini #undef __FUNCT__ 6183b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 618408122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6185b96c3477SStefano Zampini { 6186a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6187b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6188b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6189a64f4aa4SStefano Zampini Mat S_j; 6190b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6191b96c3477SStefano Zampini PetscBool free_used_adj; 6192b96c3477SStefano Zampini PetscErrorCode ierr; 6193b96c3477SStefano Zampini 6194b96c3477SStefano Zampini PetscFunctionBegin; 6195b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6196b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 619708122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6198b96c3477SStefano Zampini used_xadj = NULL; 6199b96c3477SStefano Zampini used_adjncy = NULL; 6200b96c3477SStefano Zampini } else { 620108122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 620208122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 620308122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 620408122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6205b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6206b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6207b96c3477SStefano Zampini } else { 62082fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6209b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6210b96c3477SStefano Zampini PetscInt nvtxs; 6211b96c3477SStefano Zampini 62122fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 62132fffb893SStefano Zampini if (flg_row) { 6214b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6215b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6216b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6217b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62182fffb893SStefano Zampini } else { 62192fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62202fffb893SStefano Zampini used_xadj = NULL; 62212fffb893SStefano Zampini used_adjncy = NULL; 62222fffb893SStefano Zampini } 62232fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6224b96c3477SStefano Zampini } 6225b96c3477SStefano Zampini } 6226d5574798SStefano Zampini 6227d5574798SStefano Zampini /* setup sub_schurs data */ 6228a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6229df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6230df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6231a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6232*9d54b7f4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,NULL,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 6233a64f4aa4SStefano Zampini } else { 62346816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 623504708bb6SStefano Zampini PetscBool isseqaij; 6236a3df083aSStefano Zampini PetscInt benign_n; 623772b8c272SStefano Zampini Mat change = NULL; 6238*9d54b7f4SStefano Zampini Vec scaling = NULL; 623972b8c272SStefano Zampini IS change_primal = NULL; 6240a3df083aSStefano Zampini 62415feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62425feab87aSStefano Zampini PetscInt n_vertices; 62435feab87aSStefano Zampini 62445feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62452034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62465feab87aSStefano Zampini } 624704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 624804708bb6SStefano Zampini if (!isseqaij) { 624904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 625004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 625104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 625204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 625304708bb6SStefano Zampini } else { 6254511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 625504708bb6SStefano Zampini } 625604708bb6SStefano Zampini } 6257a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6258a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6259ca92afb2SStefano Zampini } else { 6260a3df083aSStefano Zampini benign_n = 0; 6261ca92afb2SStefano Zampini } 6262bf3a8328SStefano Zampini /* If the user defines additional constraints, we import them here. 6263bf3a8328SStefano 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 */ 626472b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 626588c03ad3SStefano Zampini PC_IS *pcisf; 626688c03ad3SStefano Zampini PC_BDDC *pcbddcf; 626788c03ad3SStefano Zampini PC pcf; 626888c03ad3SStefano Zampini 626988c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 627088c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 627188c03ad3SStefano Zampini } 627288c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 627388c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 627488c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 627588c03ad3SStefano Zampini /* hacks */ 627688c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 627772b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 627872b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 627972b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 628072b8c272SStefano Zampini pcisf->n = pcis->n; 628172b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 628288c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 628388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 628488c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 628588c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 628688c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 628788c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 628872b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 628988c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 629088c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 629172b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 629272b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 629372b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 629472b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 629572b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 629688c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 629772b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 629888c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 629988c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 630088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 630188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 630288c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 630388c03ad3SStefano Zampini pcf->ops->destroy = NULL; 630488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 630588c03ad3SStefano Zampini } 6306*9d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 6307*9d54b7f4SStefano 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,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 630872b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 630972b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6310ca92afb2SStefano Zampini } 6311d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6312b96c3477SStefano Zampini 6313b96c3477SStefano Zampini /* free adjacency */ 6314b96c3477SStefano Zampini if (free_used_adj) { 6315b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6316b96c3477SStefano Zampini } 6317b96c3477SStefano Zampini PetscFunctionReturn(0); 6318b96c3477SStefano Zampini } 6319b96c3477SStefano Zampini 6320b96c3477SStefano Zampini #undef __FUNCT__ 6321b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 632208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6323b96c3477SStefano Zampini { 6324b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6325b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6326b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6327b96c3477SStefano Zampini PCBDDCGraph graph; 6328b96c3477SStefano Zampini PetscErrorCode ierr; 6329b96c3477SStefano Zampini 6330b96c3477SStefano Zampini PetscFunctionBegin; 6331b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 633208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63333301b35fSStefano Zampini IS verticesIS,verticescomm; 63343301b35fSStefano Zampini PetscInt vsize,*idxs; 6335b96c3477SStefano Zampini 6336b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63373301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63383301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63393301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63403301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63413301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6342b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63437fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6344441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63453301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6346b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6347b96c3477SStefano Zampini /* 6348b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6349b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6350b96c3477SStefano Zampini } 6351b96c3477SStefano Zampini */ 6352b96c3477SStefano Zampini } else { 6353b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6354b96c3477SStefano Zampini } 6355b96c3477SStefano Zampini 6356b96c3477SStefano Zampini /* sub_schurs init */ 6357a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6358a64f4aa4SStefano Zampini 6359b96c3477SStefano Zampini /* free graph struct */ 636008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6361b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6362b96c3477SStefano Zampini } 6363b96c3477SStefano Zampini PetscFunctionReturn(0); 6364b96c3477SStefano Zampini } 6365fa34dd3eSStefano Zampini 6366fa34dd3eSStefano Zampini #undef __FUNCT__ 6367fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6368fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6369fa34dd3eSStefano Zampini { 6370fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6371fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6372fa34dd3eSStefano Zampini PetscErrorCode ierr; 6373fa34dd3eSStefano Zampini 6374fa34dd3eSStefano Zampini PetscFunctionBegin; 6375fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6376fa34dd3eSStefano Zampini IS zerodiag = NULL; 63774f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6378fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 63794f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 638075c01103SStefano Zampini PetscReal norm; 6381fa34dd3eSStefano Zampini PetscInt i; 6382fa34dd3eSStefano Zampini 6383fa34dd3eSStefano Zampini /* B0 and B0_B */ 6384fa34dd3eSStefano Zampini if (zerodiag) { 6385fa34dd3eSStefano Zampini IS dummy; 6386fa34dd3eSStefano Zampini 63874f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 63884f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6389fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6390fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6391fa34dd3eSStefano Zampini } 6392fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6393fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6394fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6395fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6396fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6397fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6398fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6399fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6400fa34dd3eSStefano Zampini /* S_j */ 6401fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6402fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6403fa34dd3eSStefano Zampini 6404fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6405fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6406fa34dd3eSStefano Zampini /* continuous in primal space */ 6407fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6408fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6409fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6410fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 64114f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 64124f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6413fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6414fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6415fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6416fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6417fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6418fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6419fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6420fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6421fa34dd3eSStefano Zampini 6422fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6423fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6424fa34dd3eSStefano Zampini /* local with Schur */ 6425fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6426fa34dd3eSStefano Zampini if (zerodiag) { 6427fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64284f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6429fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6430fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6431fa34dd3eSStefano Zampini } 6432fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6433fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6434fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6435fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6436fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6437fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6438fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6439fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6440fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6441fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6442fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6443fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6444fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6445fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6446fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6447fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6448fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6449fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6450fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6451fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6452fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6453fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6454fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6455fa34dd3eSStefano Zampini if (zerodiag) { 6456fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6457fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64584f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6459fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6460fa34dd3eSStefano Zampini } 6461fa34dd3eSStefano Zampini /* BDDC */ 6462fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6463fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6464fa34dd3eSStefano Zampini 6465fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6466fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6467fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6468fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64694f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64704f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6471fa34dd3eSStefano Zampini } 64724f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6473fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6474fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6475fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6476fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6477fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6478fa34dd3eSStefano Zampini } 6479fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6480fa34dd3eSStefano Zampini } 6481