11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5daf8a457SStefano Zampini #include <petsc/private/sfimpl.h> 6674ae819SStefano Zampini 7a3df083aSStefano Zampini #undef __FUNCT__ 8a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 9a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 10a3df083aSStefano Zampini { 11a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 12a3df083aSStefano Zampini PetscErrorCode ierr; 13a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 14a3df083aSStefano Zampini 15a3df083aSStefano Zampini PetscFunctionBegin; 16a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 17a3df083aSStefano Zampini if (transpose) { 18a3df083aSStefano Zampini apply_right = ctx->apply_left; 19a3df083aSStefano Zampini apply_left = ctx->apply_right; 20a3df083aSStefano Zampini } else { 21a3df083aSStefano Zampini apply_right = ctx->apply_right; 22a3df083aSStefano Zampini apply_left = ctx->apply_left; 23a3df083aSStefano Zampini } 24a3df083aSStefano Zampini reset_x = PETSC_FALSE; 25a3df083aSStefano Zampini if (apply_right) { 26a3df083aSStefano Zampini const PetscScalar *ax; 27a3df083aSStefano Zampini PetscInt nl,i; 28a3df083aSStefano Zampini 29a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 30a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 31a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 32a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 33a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 34a3df083aSStefano Zampini PetscScalar sum,val; 35a3df083aSStefano Zampini const PetscInt *idxs; 36a3df083aSStefano Zampini PetscInt nz,j; 37a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 38a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 39a3df083aSStefano Zampini sum = 0.; 40a3df083aSStefano Zampini if (ctx->apply_p0) { 41a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 42a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 43a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 44a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 45a3df083aSStefano Zampini } 46a3df083aSStefano Zampini } else { 47a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 48a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 49a3df083aSStefano Zampini } 50a3df083aSStefano Zampini } 51a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 52a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 53a3df083aSStefano Zampini } 54a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 55a3df083aSStefano Zampini reset_x = PETSC_TRUE; 56a3df083aSStefano Zampini } 57a3df083aSStefano Zampini if (transpose) { 58a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 59a3df083aSStefano Zampini } else { 60a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 61a3df083aSStefano Zampini } 62a3df083aSStefano Zampini if (reset_x) { 63a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 64a3df083aSStefano Zampini } 65a3df083aSStefano Zampini if (apply_left) { 66a3df083aSStefano Zampini PetscScalar *ay; 67a3df083aSStefano Zampini PetscInt i; 68a3df083aSStefano Zampini 69a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 70a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 71a3df083aSStefano Zampini PetscScalar sum,val; 72a3df083aSStefano Zampini const PetscInt *idxs; 73a3df083aSStefano Zampini PetscInt nz,j; 74a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 75a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 76a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 77a3df083aSStefano Zampini if (ctx->apply_p0) { 78a3df083aSStefano Zampini sum = 0.; 79a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 80a3df083aSStefano Zampini sum += ay[idxs[j]]; 81a3df083aSStefano Zampini ay[idxs[j]] += val; 82a3df083aSStefano Zampini } 83a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 84a3df083aSStefano Zampini } else { 85a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 86a3df083aSStefano Zampini ay[idxs[j]] += val; 87a3df083aSStefano Zampini } 88a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 89a3df083aSStefano Zampini } 90a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 91a3df083aSStefano Zampini } 92a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 93a3df083aSStefano Zampini } 94a3df083aSStefano Zampini PetscFunctionReturn(0); 95a3df083aSStefano Zampini } 96a3df083aSStefano Zampini 97a3df083aSStefano Zampini #undef __FUNCT__ 98a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 99a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 100a3df083aSStefano Zampini { 101a3df083aSStefano Zampini PetscErrorCode ierr; 102a3df083aSStefano Zampini 103a3df083aSStefano Zampini PetscFunctionBegin; 104a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 105a3df083aSStefano Zampini PetscFunctionReturn(0); 106a3df083aSStefano Zampini } 107a3df083aSStefano Zampini 108a3df083aSStefano Zampini #undef __FUNCT__ 109a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 110a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 111a3df083aSStefano Zampini { 112a3df083aSStefano Zampini PetscErrorCode ierr; 113a3df083aSStefano Zampini 114a3df083aSStefano Zampini PetscFunctionBegin; 115a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 116a3df083aSStefano Zampini PetscFunctionReturn(0); 117a3df083aSStefano Zampini } 118a3df083aSStefano Zampini 119a3df083aSStefano Zampini #undef __FUNCT__ 120a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 121a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 122a3df083aSStefano Zampini { 123a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 124a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 125a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 126a3df083aSStefano Zampini PetscErrorCode ierr; 127a3df083aSStefano Zampini 128a3df083aSStefano Zampini PetscFunctionBegin; 129a3df083aSStefano Zampini if (!restore) { 1301dd7afcfSStefano Zampini Mat A_IB,A_BI; 131a3df083aSStefano Zampini PetscScalar *work; 132df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 133a3df083aSStefano Zampini 1341dd7afcfSStefano Zampini if (pcbddc->benign_original_mat) { 1351dd7afcfSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 1361dd7afcfSStefano Zampini } 1371dd7afcfSStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) { 1381dd7afcfSStefano Zampini PetscFunctionReturn(0); 1391dd7afcfSStefano Zampini } 140a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 141a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 142a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 145a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 146a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 147a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 148a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 149a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 150a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 151a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 152059032f7SStefano Zampini if (reuse) { 153a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 1541dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 155059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 156059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 157059032f7SStefano Zampini PetscInt i; 158059032f7SStefano Zampini 159059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 160059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 161059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 162059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 163059032f7SStefano Zampini } 164059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 1651dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 166059032f7SStefano Zampini } 167a3df083aSStefano Zampini ctx->A = pcis->A_IB; 168a3df083aSStefano Zampini ctx->work = work; 169a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 170a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 171a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 172a3df083aSStefano Zampini pcis->A_IB = A_IB; 173a3df083aSStefano Zampini 174a3df083aSStefano Zampini /* A_BI as A_IB^T */ 175a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 176a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 177a3df083aSStefano Zampini pcis->A_BI = A_BI; 178a3df083aSStefano Zampini } else { 1791dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 1801dd7afcfSStefano Zampini PetscFunctionReturn(0); 1811dd7afcfSStefano Zampini } 182a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 183a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 184a3df083aSStefano Zampini pcis->A_IB = ctx->A; 1851dd7afcfSStefano Zampini ctx->A = NULL; 1861dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 1871dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 1881dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 1891dd7afcfSStefano Zampini if (ctx->free) { 190059032f7SStefano Zampini PetscInt i; 1911dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 192059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 193059032f7SStefano Zampini } 194059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 195059032f7SStefano Zampini } 196a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 197a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 198a3df083aSStefano Zampini } 199a3df083aSStefano Zampini PetscFunctionReturn(0); 200a3df083aSStefano Zampini } 201a3df083aSStefano Zampini 202a3df083aSStefano Zampini /* used just in bddc debug mode */ 203a3df083aSStefano Zampini #undef __FUNCT__ 204a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 205a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 206a3df083aSStefano Zampini { 207a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 208a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 209a3df083aSStefano Zampini Mat An; 210a3df083aSStefano Zampini PetscErrorCode ierr; 211a3df083aSStefano Zampini 212a3df083aSStefano Zampini PetscFunctionBegin; 213a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 214a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 215a3df083aSStefano Zampini if (is1) { 216a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 217a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 218a3df083aSStefano Zampini } else { 219a3df083aSStefano Zampini *B = An; 220a3df083aSStefano Zampini } 221a3df083aSStefano Zampini PetscFunctionReturn(0); 222a3df083aSStefano Zampini } 223a3df083aSStefano Zampini 2241cf9b237SStefano Zampini /* TODO: add reuse flag */ 2251cf9b237SStefano Zampini #undef __FUNCT__ 2261cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2271cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2281cf9b237SStefano Zampini { 2291cf9b237SStefano Zampini Mat Bt; 2301cf9b237SStefano Zampini PetscScalar *a,*bdata; 2311cf9b237SStefano Zampini const PetscInt *ii,*ij; 2321cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2331cf9b237SStefano Zampini PetscBool flg_row; 2341cf9b237SStefano Zampini PetscErrorCode ierr; 2351cf9b237SStefano Zampini 2361cf9b237SStefano Zampini PetscFunctionBegin; 2371cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2381cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2391cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2401cf9b237SStefano Zampini nnz = n; 2411cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2421cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2431cf9b237SStefano Zampini } 2441cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2451cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2461cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2471cf9b237SStefano Zampini nnz = 0; 2481cf9b237SStefano Zampini bii[0] = 0; 2491cf9b237SStefano Zampini for (i=0;i<n;i++) { 2501cf9b237SStefano Zampini PetscInt j; 2511cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2521cf9b237SStefano Zampini PetscScalar entry = a[j]; 2531cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2541cf9b237SStefano Zampini bij[nnz] = ij[j]; 2551cf9b237SStefano Zampini bdata[nnz] = entry; 2561cf9b237SStefano Zampini nnz++; 2571cf9b237SStefano Zampini } 2581cf9b237SStefano Zampini } 2591cf9b237SStefano Zampini bii[i+1] = nnz; 2601cf9b237SStefano Zampini } 2611cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2621cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2631cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2641cf9b237SStefano Zampini { 2651cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2661cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2671cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2681cf9b237SStefano Zampini } 2691cf9b237SStefano Zampini *B = Bt; 2701cf9b237SStefano Zampini PetscFunctionReturn(0); 2711cf9b237SStefano Zampini } 2721cf9b237SStefano Zampini 273674ae819SStefano Zampini #undef __FUNCT__ 2744f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2754f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2764f1b2e48SStefano Zampini { 2774f1b2e48SStefano Zampini Mat B; 2784f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2794f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2804f1b2e48SStefano Zampini PCBDDCGraph graph; 2814f1b2e48SStefano Zampini PetscInt i,n; 2824f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 2834f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 2844f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 2854f1b2e48SStefano Zampini PetscErrorCode ierr; 2864f1b2e48SStefano Zampini 2874f1b2e48SStefano Zampini PetscFunctionBegin; 28863c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 28963c961adSStefano Zampini *ncc = 0; 29063c961adSStefano Zampini *cc = NULL; 29163c961adSStefano Zampini PetscFunctionReturn(0); 29263c961adSStefano Zampini } 2934f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 2944f1b2e48SStefano Zampini if (!isseqaij && filter) { 2951cf9b237SStefano Zampini PetscBool isseqdense; 2961cf9b237SStefano Zampini 2971cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 2981cf9b237SStefano Zampini if (!isseqdense) { 2994f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3001cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3011cf9b237SStefano Zampini PetscScalar *array; 3021cf9b237SStefano Zampini PetscReal chop=1.e-6; 3031cf9b237SStefano Zampini 3041cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3051cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3061cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3071cf9b237SStefano Zampini for (i=0;i<n;i++) { 3081cf9b237SStefano Zampini PetscInt j; 3091cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3101cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3111cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3121cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3131cf9b237SStefano Zampini } 3141cf9b237SStefano Zampini } 3151cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3169d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 3171cf9b237SStefano Zampini } 3184f1b2e48SStefano Zampini } else { 3194f1b2e48SStefano Zampini B = A; 3204f1b2e48SStefano Zampini } 3214f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3224f1b2e48SStefano Zampini 3234f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3244f1b2e48SStefano Zampini if (filter) { 3254f1b2e48SStefano Zampini PetscScalar *data; 3264f1b2e48SStefano Zampini PetscInt j,cum; 3274f1b2e48SStefano Zampini 3284f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3294f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3304f1b2e48SStefano Zampini cum = 0; 3314f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3324f1b2e48SStefano Zampini PetscInt t; 3334f1b2e48SStefano Zampini 3344f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3354f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3364f1b2e48SStefano Zampini continue; 3374f1b2e48SStefano Zampini } 3384f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3394f1b2e48SStefano Zampini } 3404f1b2e48SStefano Zampini t = xadj_filtered[i]; 3414f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3424f1b2e48SStefano Zampini cum += t; 3434f1b2e48SStefano Zampini } 3444f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3454f1b2e48SStefano Zampini } else { 3464f1b2e48SStefano Zampini xadj_filtered = NULL; 3474f1b2e48SStefano Zampini adjncy_filtered = NULL; 3484f1b2e48SStefano Zampini } 3494f1b2e48SStefano Zampini 3504f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3514f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3524f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3534f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3544f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3554f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3564f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3574f1b2e48SStefano Zampini if (xadj_filtered) { 3584f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3594f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3604f1b2e48SStefano Zampini } else { 3614f1b2e48SStefano Zampini graph->xadj = xadj; 3624f1b2e48SStefano Zampini graph->adjncy = adjncy; 3634f1b2e48SStefano Zampini } 3644f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3654f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3664f1b2e48SStefano Zampini /* partial clean up */ 3674f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3684f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3691cf9b237SStefano Zampini if (A != B) { 3704f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3714f1b2e48SStefano Zampini } 3724f1b2e48SStefano Zampini 3734f1b2e48SStefano Zampini /* get back data */ 3741cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3751cf9b237SStefano Zampini if (cc) { 3764f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3774f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3784f1b2e48SStefano 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); 3794f1b2e48SStefano Zampini } 3804f1b2e48SStefano Zampini *cc = cc_n; 3811cf9b237SStefano Zampini } 3824f1b2e48SStefano Zampini /* clean up graph */ 3834f1b2e48SStefano Zampini graph->xadj = 0; 3844f1b2e48SStefano Zampini graph->adjncy = 0; 3854f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 3864f1b2e48SStefano Zampini PetscFunctionReturn(0); 3874f1b2e48SStefano Zampini } 3884f1b2e48SStefano Zampini 3894f1b2e48SStefano Zampini #undef __FUNCT__ 3905408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 3915408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 3925408967cSStefano Zampini { 3935408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3945408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 395dee84bffSStefano Zampini IS dirIS = NULL; 3964f1b2e48SStefano Zampini PetscInt i; 3975408967cSStefano Zampini PetscErrorCode ierr; 3985408967cSStefano Zampini 3995408967cSStefano Zampini PetscFunctionBegin; 400dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4015408967cSStefano Zampini if (zerodiag) { 4025408967cSStefano Zampini Mat A; 4035408967cSStefano Zampini Vec vec3_N; 4045408967cSStefano Zampini PetscScalar *vals; 4055408967cSStefano Zampini const PetscInt *idxs; 406d12d3064SStefano Zampini PetscInt nz,*count; 4075408967cSStefano Zampini 4085408967cSStefano Zampini /* p0 */ 4095408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4105408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4115408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4125408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4134f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4145408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4155408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4165408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4175408967cSStefano Zampini /* v_I */ 4185408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4195408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4205408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4215408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4225408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4235408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4245408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4255408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4265408967cSStefano Zampini if (dirIS) { 4275408967cSStefano Zampini PetscInt n; 4285408967cSStefano Zampini 4295408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4305408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4315408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4325408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4335408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4345408967cSStefano Zampini } 4355408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4365408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4375408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4385408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4415408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 442fbfcb133SStefano Zampini if (PetscAbsScalar(vals[0]) > 1.e-1) { 443b9b0e38cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %1.6e (should be numerically 0.)",PetscAbsScalar(vals[0])); 4445408967cSStefano Zampini } 4455408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4465408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 447d12d3064SStefano Zampini 448d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 449d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 450d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 451d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 452d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 453d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 454d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 455d12d3064SStefano Zampini if (count[idxs[i]]) { 456d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 457d12d3064SStefano Zampini } 458d4d8cf7bSStefano Zampini } 459d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 460d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4615408967cSStefano Zampini } 462dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4635408967cSStefano Zampini 4645408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4655408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4664f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4675408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4684f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4695408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4704f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4714f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4724f1b2e48SStefano 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); 4734f1b2e48SStefano Zampini } 4745408967cSStefano Zampini } 4755408967cSStefano Zampini PetscFunctionReturn(0); 4765408967cSStefano Zampini } 4775408967cSStefano Zampini 4785408967cSStefano Zampini #undef __FUNCT__ 479339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 480339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 481339f8db1SStefano Zampini { 482339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4834f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 484b0f5fe93SStefano Zampini PetscInt nz,n; 4854f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 486339f8db1SStefano Zampini PetscErrorCode ierr; 487339f8db1SStefano Zampini 488339f8db1SStefano Zampini PetscFunctionBegin; 4899f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4909f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 491a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 492a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 493a3df083aSStefano Zampini } 494a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 495a3df083aSStefano Zampini pcbddc->benign_n = 0; 4964f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4974f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 4984f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 4994f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 5004f1b2e48SStefano Zampini since the local Schur complements are SPD 5014f1b2e48SStefano Zampini */ 5024f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5034f1b2e48SStefano Zampini have_null = PETSC_TRUE; 50440fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5054f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5064f1b2e48SStefano Zampini 5074f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5084f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5094f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5104f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 511ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 51240fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 51340fa8d13SStefano Zampini if (!sorted) { 51440fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 51540fa8d13SStefano Zampini } 51640fa8d13SStefano Zampini } else { 51740fa8d13SStefano Zampini pressures = NULL; 51840fa8d13SStefano Zampini } 51997d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 52097d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 52127b6a85dSStefano Zampini if (!n) pcbddc->benign_change_explicit = PETSC_TRUE; 52297d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 523339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 524339f8db1SStefano Zampini if (!sorted) { 525339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 526339f8db1SStefano Zampini } 527339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5284f1b2e48SStefano Zampini if (!nz) { 5294f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5304f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 53140fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 53240fa8d13SStefano Zampini } 5334f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5344f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5354f1b2e48SStefano Zampini zerodiag_subs = NULL; 5364f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5374f1b2e48SStefano Zampini if (has_null_pressures) { 5384f1b2e48SStefano Zampini IS *subs; 5394f1b2e48SStefano Zampini PetscInt nsubs,i; 5404f1b2e48SStefano Zampini 5414f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5424f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5434f1b2e48SStefano Zampini if (nsubs > 1) { 5444f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5454f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5464f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5474f1b2e48SStefano Zampini IS t_zerodiag_subs; 5484f1b2e48SStefano Zampini PetscInt nl; 5494f1b2e48SStefano Zampini 5504f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5514f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5524f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5534f1b2e48SStefano Zampini if (nl) { 5544f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5554f1b2e48SStefano Zampini 5564f1b2e48SStefano Zampini if (pressures) { 5574f1b2e48SStefano Zampini IS t_pressure_subs; 5584f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5594f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5604f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5614f1b2e48SStefano Zampini } 5624f1b2e48SStefano Zampini if (valid) { 5634f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5644f1b2e48SStefano Zampini pcbddc->benign_n++; 5654f1b2e48SStefano Zampini } else { 5664f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5674f1b2e48SStefano Zampini } 5684f1b2e48SStefano Zampini } 5694f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5704f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5714f1b2e48SStefano Zampini } 5724f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5734f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5744f1b2e48SStefano Zampini if (pressures) { 5754f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5764f1b2e48SStefano Zampini } 5774f1b2e48SStefano Zampini if (valid) { 5784f1b2e48SStefano Zampini pcbddc->benign_n = 1; 579ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5804f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5814f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5824f1b2e48SStefano Zampini } 5834f1b2e48SStefano Zampini } 5844f1b2e48SStefano Zampini } 5854f1b2e48SStefano Zampini 5864f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 587b9b0e38cSStefano Zampini PetscInt n; 588b9b0e38cSStefano Zampini 5894f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5904f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 591b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 592b9b0e38cSStefano Zampini if (n) { 5934f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5944f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5954f1b2e48SStefano Zampini } 596b9b0e38cSStefano Zampini } 5974f1b2e48SStefano Zampini 5984f1b2e48SStefano Zampini /* final check for null pressures */ 5994f1b2e48SStefano Zampini if (zerodiag && pressures) { 6004f1b2e48SStefano Zampini PetscInt nz,np; 6014f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6024f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 6044f1b2e48SStefano Zampini } 6054f1b2e48SStefano Zampini 6064f1b2e48SStefano Zampini if (recompute_zerodiag) { 6074f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6084f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6094f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6104f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6114f1b2e48SStefano Zampini } else { 6124f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6134f1b2e48SStefano Zampini 6144f1b2e48SStefano Zampini nzn = 0; 6154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6164f1b2e48SStefano Zampini PetscInt ns; 6174f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6184f1b2e48SStefano Zampini nzn += ns; 6194f1b2e48SStefano Zampini } 6204f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini nzn = 0; 6224f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6234f1b2e48SStefano Zampini PetscInt ns,*idxs; 6244f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6254f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6264f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6274f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6284f1b2e48SStefano Zampini nzn += ns; 6294f1b2e48SStefano Zampini } 6304f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6314f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6324f1b2e48SStefano Zampini } 6334f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6344f1b2e48SStefano Zampini } 6354f1b2e48SStefano Zampini 636b3afcdbeSStefano Zampini /* no-net-flux */ 637b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 638b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 639b3afcdbeSStefano Zampini MatNullSpace near_null_space; 640b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 641b3afcdbeSStefano Zampini Vec quad_vec; 642b3afcdbeSStefano Zampini PetscScalar *pvals; 643b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 644b3afcdbeSStefano Zampini 645b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 646b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 647b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 648b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 649b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 650b3afcdbeSStefano Zampini 651b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 652b3afcdbeSStefano Zampini np = 0; 653b3afcdbeSStefano Zampini if (zerodiag) { 654b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 655b3afcdbeSStefano Zampini } else if (pressures) { 656b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 657b3afcdbeSStefano Zampini } 658b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 659b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 660b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 661b3afcdbeSStefano Zampini if (np) { 662b3afcdbeSStefano Zampini const PetscInt *pidxs; 663b3afcdbeSStefano Zampini 664b3afcdbeSStefano Zampini if (zerodiag) { 665b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 666b3afcdbeSStefano Zampini } else if (pressures) { 667b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 668b3afcdbeSStefano Zampini } 669b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 670b3afcdbeSStefano Zampini if (zerodiag) { 671b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 672b3afcdbeSStefano Zampini } else if (pressures) { 673b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 674b3afcdbeSStefano Zampini } 675b3afcdbeSStefano Zampini } 676b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 677b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 678b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 679b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 680b3afcdbeSStefano Zampini 681b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 682b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 683daf8a457SStefano Zampini for (i=0;i<matis->sf->nroots;i++) matis->sf_rootdata[i] = -1; 684daf8a457SStefano Zampini for (i=0;i<matis->sf->nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 685b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 686b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 687b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 688b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 689b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 690daf8a457SStefano Zampini for (i=0;i<matis->sf->nleaves;i++) { 691b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 692b3afcdbeSStefano Zampini pvals[i] = 0.; 693b3afcdbeSStefano Zampini } 694b3afcdbeSStefano Zampini } 695daf8a457SStefano Zampini ierr = PetscMalloc1(matis->sf->nleaves,&dummyins);CHKERRQ(ierr); 696daf8a457SStefano Zampini for (i=0;i<matis->sf->nleaves;i++) dummyins[i] = i; 697daf8a457SStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf->nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 698b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 699b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 700b3afcdbeSStefano Zampini 701b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 702b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 703b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 704b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 705b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 706b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 707b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 708b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 709b3afcdbeSStefano Zampini } 710b3afcdbeSStefano Zampini 711b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7124f1b2e48SStefano Zampini if (has_null_pressures) { 7134f1b2e48SStefano Zampini IS zerodiagc; 7144f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7154f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7164f1b2e48SStefano Zampini 7174f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 718339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 719339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 720339f8db1SStefano Zampini /* local change of basis for pressures */ 721339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 72297d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 723339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 724339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 725339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7264f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7274f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7284f1b2e48SStefano Zampini PetscInt nzs,j; 7294f1b2e48SStefano Zampini 7304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7314f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7324f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7334f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7344f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7354f1b2e48SStefano Zampini } 736339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 737339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 738339f8db1SStefano Zampini /* set identity on velocities */ 739339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 740339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 741339f8db1SStefano Zampini } 7424f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7434f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7449f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7454f1b2e48SStefano 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); 746339f8db1SStefano Zampini /* set change on pressures */ 7474f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7484f1b2e48SStefano Zampini PetscScalar *array; 7494f1b2e48SStefano Zampini PetscInt nzs; 7504f1b2e48SStefano Zampini 7514f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7524f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7534f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 754339f8db1SStefano Zampini PetscScalar vals[2]; 755339f8db1SStefano Zampini PetscInt cols[2]; 756339f8db1SStefano Zampini 757339f8db1SStefano Zampini cols[0] = idxs[i]; 7584f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 759339f8db1SStefano Zampini vals[0] = 1.; 760b0f5fe93SStefano Zampini vals[1] = 1.; 7614f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 762339f8db1SStefano Zampini } 7634f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7644f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7654f1b2e48SStefano Zampini array[nzs-1] = 1.; 7664f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7674f1b2e48SStefano Zampini /* store local idxs for p0 */ 7684f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7694f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 770339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7714f1b2e48SStefano Zampini } 772339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 773339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 774a3df083aSStefano Zampini /* project if needed */ 775a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7761dd7afcfSStefano Zampini Mat M; 7771dd7afcfSStefano Zampini 7781dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 779339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 7801dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 7811dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 782a3df083aSStefano Zampini } 7834f1b2e48SStefano Zampini /* store global idxs for p0 */ 7844f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 785339f8db1SStefano Zampini } 786ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7874f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 788b0f5fe93SStefano Zampini 789b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 790b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 79127b6a85dSStefano Zampini /* determines if the problem has subdomains with 0 pressure block */ 79227b6a85dSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 793339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 794339f8db1SStefano Zampini PetscFunctionReturn(0); 795339f8db1SStefano Zampini } 796339f8db1SStefano Zampini 797339f8db1SStefano Zampini #undef __FUNCT__ 798015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 799015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 800efc2fbd9SStefano Zampini { 801efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 802de9d7bd0SStefano Zampini PetscScalar *array; 803efc2fbd9SStefano Zampini PetscErrorCode ierr; 804efc2fbd9SStefano Zampini 805efc2fbd9SStefano Zampini PetscFunctionBegin; 806efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 807efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 8084f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 809efc2fbd9SStefano Zampini } 810de9d7bd0SStefano Zampini if (get) { 811efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8124f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8134f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 814efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 815de9d7bd0SStefano Zampini } else { 816de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 817de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 818de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 819de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 820efc2fbd9SStefano Zampini } 821efc2fbd9SStefano Zampini PetscFunctionReturn(0); 822efc2fbd9SStefano Zampini } 823efc2fbd9SStefano Zampini 824efc2fbd9SStefano Zampini #undef __FUNCT__ 825c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 826c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 827c263805aSStefano Zampini { 828c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 829c263805aSStefano Zampini PetscErrorCode ierr; 830c263805aSStefano Zampini 831c263805aSStefano Zampini PetscFunctionBegin; 832c263805aSStefano Zampini /* TODO: add error checking 833c263805aSStefano Zampini - avoid nested pop (or push) calls. 834c263805aSStefano Zampini - cannot push before pop. 8351c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 836c263805aSStefano Zampini */ 8374f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 838efc2fbd9SStefano Zampini PetscFunctionReturn(0); 839efc2fbd9SStefano Zampini } 840c263805aSStefano Zampini if (pop) { 841a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8424f1b2e48SStefano Zampini IS is_p0; 8434f1b2e48SStefano Zampini MatReuse reuse; 844c263805aSStefano Zampini 845c263805aSStefano Zampini /* extract B_0 */ 8464f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8474f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8484f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8494f1b2e48SStefano Zampini } 8504f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8514f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 852c263805aSStefano Zampini /* remove rows and cols from local problem */ 853c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 85497d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8554f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8564f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 857a3df083aSStefano Zampini } else { 858a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 859a3df083aSStefano Zampini PetscScalar *vals; 860a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 861a3df083aSStefano Zampini 862a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 863a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 864a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8650b5adadeSStefano Zampini PetscInt *nnz; 866a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 867a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 868a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 869331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 870331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 871331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 872331e053bSStefano Zampini nnz[i] = n - nnz[i]; 873331e053bSStefano Zampini } 874331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 875331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 876331e053bSStefano Zampini } 877a3df083aSStefano Zampini 878a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 879a3df083aSStefano Zampini PetscScalar *array; 880a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 881a3df083aSStefano Zampini 882a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 883a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 884a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 885a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 886a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 887a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 888a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 889a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 890a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 891a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 892a3df083aSStefano Zampini cum = 0; 893a3df083aSStefano Zampini for (j=0;j<n;j++) { 89422db5ddcSStefano Zampini if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) { 895a3df083aSStefano Zampini vals[cum] = array[j]; 896a3df083aSStefano Zampini idxs_ins[cum] = j; 897a3df083aSStefano Zampini cum++; 898a3df083aSStefano Zampini } 899a3df083aSStefano Zampini } 900a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 901a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 902a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 903a3df083aSStefano Zampini } 904a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 905a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 906a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 907a3df083aSStefano Zampini } 908c263805aSStefano Zampini } else { /* push */ 909a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9104f1b2e48SStefano Zampini PetscInt i; 9114f1b2e48SStefano Zampini 9124f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9134f1b2e48SStefano Zampini PetscScalar *B0_vals; 9144f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9154f1b2e48SStefano Zampini 9164f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9174f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9187b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9194f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9204f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9214f1b2e48SStefano Zampini } 922c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 923c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 924a3df083aSStefano Zampini } else { 925a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 926a3df083aSStefano Zampini } 927c263805aSStefano Zampini } 928c263805aSStefano Zampini PetscFunctionReturn(0); 929c263805aSStefano Zampini } 930c263805aSStefano Zampini 931c263805aSStefano Zampini #undef __FUNCT__ 932b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 93308122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 934b1b3d7a2SStefano Zampini { 935b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 93608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 93708122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 93808122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 93908122e43SStefano Zampini PetscScalar *work,lwork; 94008122e43SStefano Zampini PetscScalar *St,*S,*eigv; 94108122e43SStefano Zampini PetscScalar *Sarray,*Starray; 94208122e43SStefano Zampini PetscReal *eigs,thresh; 9431b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 944f6f667cfSStefano Zampini PetscBool allocated_S_St; 94508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 94608122e43SStefano Zampini PetscReal *rwork; 94708122e43SStefano Zampini #endif 948b1b3d7a2SStefano Zampini PetscErrorCode ierr; 949b1b3d7a2SStefano Zampini 950b1b3d7a2SStefano Zampini PetscFunctionBegin; 951af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 952af25d912SStefano 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); 95306a4e24aSStefano Zampini 954fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 955fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 956fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 957fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9581575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 959fd14bc51SStefano Zampini } 960fd14bc51SStefano Zampini 961e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 962e496cd5dSStefano 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); 963e496cd5dSStefano Zampini } 964e496cd5dSStefano Zampini 96508122e43SStefano Zampini /* max size of subsets */ 96608122e43SStefano Zampini mss = 0; 96708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 96808122e43SStefano Zampini PetscInt subset_size; 969862806e4SStefano Zampini 97008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 97108122e43SStefano Zampini mss = PetscMax(mss,subset_size); 97208122e43SStefano Zampini } 97308122e43SStefano Zampini 97408122e43SStefano Zampini /* min/max and threshold */ 97508122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 976f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 97708122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 978f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 979f6f667cfSStefano Zampini if (nmin) { 980f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 981f6f667cfSStefano Zampini } 98208122e43SStefano Zampini 98308122e43SStefano Zampini /* allocate lapack workspace */ 98408122e43SStefano Zampini cum = cum2 = 0; 98508122e43SStefano Zampini maxneigs = 0; 98608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 98708122e43SStefano Zampini PetscInt n,subset_size; 988f6f667cfSStefano Zampini 98908122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 99008122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9919162d606SStefano Zampini cum += subset_size; 9929162d606SStefano Zampini cum2 += subset_size*n; 99308122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 99408122e43SStefano Zampini } 99508122e43SStefano Zampini if (mss) { 9969ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 99708122e43SStefano Zampini PetscBLASInt B_itype = 1; 99808122e43SStefano Zampini PetscBLASInt B_N = mss; 9994c6709b3SStefano Zampini PetscReal zero = 0.0; 10004c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 100108122e43SStefano Zampini 100208122e43SStefano Zampini B_lwork = -1; 100308122e43SStefano Zampini S = NULL; 100408122e43SStefano Zampini St = NULL; 1005a58a30b4SStefano Zampini eigs = NULL; 1006a58a30b4SStefano Zampini eigv = NULL; 1007a58a30b4SStefano Zampini B_iwork = NULL; 1008a58a30b4SStefano Zampini B_ifail = NULL; 1009d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1010d1710679SStefano Zampini rwork = NULL; 1011d1710679SStefano Zampini #endif 10128bec7fa6SStefano Zampini thresh = 1.0; 101308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 101408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 101508122e43SStefano 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)); 101608122e43SStefano Zampini #else 101708122e43SStefano 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)); 101808122e43SStefano Zampini #endif 101908122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 102008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 102108122e43SStefano Zampini } else { 102208122e43SStefano Zampini /* TODO */ 102308122e43SStefano Zampini } 102408122e43SStefano Zampini } else { 102508122e43SStefano Zampini lwork = 0; 102608122e43SStefano Zampini } 102708122e43SStefano Zampini 102808122e43SStefano Zampini nv = 0; 1029d62866d3SStefano 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) */ 1030d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 103108122e43SStefano Zampini } 10324c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1033f6f667cfSStefano Zampini if (allocated_S_St) { 1034f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1035f6f667cfSStefano Zampini } 1036f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 103708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 103808122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 103908122e43SStefano Zampini #endif 10409162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10419162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10429162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 104308122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10449162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 104508122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 104608122e43SStefano Zampini 104708122e43SStefano Zampini maxneigs = 0; 104872b8c272SStefano Zampini cum = cumarray = 0; 10499162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10509162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1051d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 105208122e43SStefano Zampini const PetscInt *idxs; 105308122e43SStefano Zampini 1054d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 105508122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 105608122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 105708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 105808122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10599162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10609162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 106108122e43SStefano Zampini } 1062d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 106308122e43SStefano Zampini } 106408122e43SStefano Zampini 106508122e43SStefano Zampini if (mss) { /* multilevel */ 106608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 106708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 106808122e43SStefano Zampini } 106908122e43SStefano Zampini 1070ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 107108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 107208122e43SStefano Zampini const PetscInt *idxs; 10739d54b7f4SStefano Zampini PetscReal upper,lower; 1074862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 107508122e43SStefano Zampini PetscBLASInt B_N; 1076aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 107708122e43SStefano Zampini 10789d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 10799d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 10809d54b7f4SStefano Zampini lower = thresh; 10819d54b7f4SStefano Zampini } else { 10829d54b7f4SStefano Zampini upper = 1./thresh; 10839d54b7f4SStefano Zampini lower = 0.; 10849d54b7f4SStefano Zampini } 1085862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1086ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1087f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1088f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10899ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1090aff50787SStefano Zampini PetscInt j,k; 1091aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1092aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1093aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109408122e43SStefano Zampini } 109508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1096aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1097aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1098aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1099aff50787SStefano Zampini } 110008122e43SStefano Zampini } 110108122e43SStefano Zampini } else { 110208122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110308122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110408122e43SStefano Zampini } 11058bec7fa6SStefano Zampini } else { 1106f6f667cfSStefano Zampini S = Sarray + cumarray; 1107f6f667cfSStefano Zampini St = Starray + cumarray; 11088bec7fa6SStefano Zampini } 1109aff50787SStefano Zampini /* see if we can save some work */ 1110b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1111aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1112aff50787SStefano Zampini } 1113aff50787SStefano Zampini 1114b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1115aff50787SStefano Zampini B_neigs = 0; 1116aff50787SStefano Zampini } else { 11179ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 111808122e43SStefano Zampini PetscBLASInt B_itype = 1; 1119f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11204c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11219552c7c7SStefano Zampini PetscInt nmin_s; 1122b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 112308122e43SStefano Zampini 1124fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11258bec7fa6SStefano 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]]); 1126fd14bc51SStefano Zampini } 1127d16cbb6bSStefano Zampini 1128b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1129b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1130b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1131b7ab4a40SStefano Zampini } 1132b7ab4a40SStefano Zampini 113308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1134b7ab4a40SStefano Zampini if (compute_range) { 1135d16cbb6bSStefano Zampini 1136d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 113708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11389d54b7f4SStefano 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)); 113908122e43SStefano Zampini #else 11409d54b7f4SStefano 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)); 114108122e43SStefano Zampini #endif 1142b7ab4a40SStefano Zampini } else if (!same_data) { 1143d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1144d16cbb6bSStefano Zampini B_IL = 1; 1145d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 11469d54b7f4SStefano 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)); 1147d16cbb6bSStefano Zampini #else 11489d54b7f4SStefano 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)); 1149d16cbb6bSStefano Zampini #endif 1150b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1151b7ab4a40SStefano Zampini PetscInt k; 1152b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1153b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1154b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1155b7ab4a40SStefano Zampini nmin = nmax; 1156b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1157b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1158b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1159b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1160b7ab4a40SStefano Zampini } 1161d16cbb6bSStefano Zampini } 116208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 116308122e43SStefano Zampini if (B_ierr) { 11646c4ed002SBarry 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); 11656c4ed002SBarry 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); 11666c4ed002SBarry 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); 116708122e43SStefano Zampini } 116808122e43SStefano Zampini 116908122e43SStefano Zampini if (B_neigs > nmax) { 1170fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1171fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1172fd14bc51SStefano Zampini } 11739d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 117408122e43SStefano Zampini B_neigs = nmax; 117508122e43SStefano Zampini } 117608122e43SStefano Zampini 11779552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11789552c7c7SStefano Zampini if (B_neigs < nmin_s) { 117908122e43SStefano Zampini PetscBLASInt B_neigs2; 118008122e43SStefano Zampini 11819d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1182f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 11839d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 11849d54b7f4SStefano Zampini } else { 11859d54b7f4SStefano Zampini B_IL = B_neigs + 1; 11869d54b7f4SStefano Zampini B_IU = nmin_s; 11879d54b7f4SStefano Zampini } 1188fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1189fd14bc51SStefano 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); 1190fd14bc51SStefano Zampini } 11919ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 119208122e43SStefano Zampini PetscInt j; 119308122e43SStefano Zampini for (j=0;j<subset_size;j++) { 119408122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 119508122e43SStefano Zampini } 119608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 119708122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 119808122e43SStefano Zampini } 119908122e43SStefano Zampini } else { 120008122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 120108122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 120208122e43SStefano Zampini } 120308122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 120408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 12059d54b7f4SStefano 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)); 120608122e43SStefano Zampini #else 12079d54b7f4SStefano 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)); 120808122e43SStefano Zampini #endif 120908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 121008122e43SStefano Zampini B_neigs += B_neigs2; 121108122e43SStefano Zampini } 121208122e43SStefano Zampini if (B_ierr) { 12136c4ed002SBarry 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); 12146c4ed002SBarry 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); 12156c4ed002SBarry 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); 121608122e43SStefano Zampini } 1217fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1218ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 121908122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 122008122e43SStefano Zampini if (eigs[j] == 0.0) { 1221ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 122208122e43SStefano Zampini } else { 12239d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1224ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 12259d54b7f4SStefano Zampini } else { 12269d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 12279d54b7f4SStefano Zampini } 1228fd14bc51SStefano Zampini } 122908122e43SStefano Zampini } 123008122e43SStefano Zampini } 123108122e43SStefano Zampini } else { 123208122e43SStefano Zampini /* TODO */ 123308122e43SStefano Zampini } 1234aff50787SStefano Zampini } 12356c3e6151SStefano Zampini /* change the basis back to the original one */ 12366c3e6151SStefano Zampini if (sub_schurs->change) { 123772b8c272SStefano Zampini Mat change,phi,phit; 12386c3e6151SStefano Zampini 12396c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12406c3e6151SStefano Zampini PetscInt ii; 12416c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12426c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12436c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12446c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12456c3e6151SStefano Zampini } 12466c3e6151SStefano Zampini } 12476c3e6151SStefano Zampini } 124872b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12496c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 125072b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12516c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12526c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12536c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12546c3e6151SStefano Zampini } 12558bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12568bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12579162d606SStefano Zampini if (B_neigs) { 12589162d606SStefano 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); 1259fd14bc51SStefano Zampini 1260fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12619552c7c7SStefano Zampini PetscInt ii; 12629552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1263ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12649552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1265ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1266ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1267ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1268ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1269ac47001eSStefano Zampini #else 1270ac47001eSStefano 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); 1271ac47001eSStefano Zampini #endif 12729552c7c7SStefano Zampini } 12739552c7c7SStefano Zampini } 1274fd14bc51SStefano Zampini } 12759162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12769162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12779162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12789162d606SStefano Zampini cum++; 127908122e43SStefano Zampini } 128008122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 128108122e43SStefano Zampini /* shift for next computation */ 128208122e43SStefano Zampini cumarray += subset_size*subset_size; 128308122e43SStefano Zampini } 1284fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1285fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1286fd14bc51SStefano Zampini } 128708122e43SStefano Zampini 128808122e43SStefano Zampini if (mss) { 128908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 129008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1291f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1292f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1293f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 129408122e43SStefano Zampini } 1295f6f667cfSStefano Zampini if (allocated_S_St) { 1296f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1297f6f667cfSStefano Zampini } 1298f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 129908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 130008122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 130108122e43SStefano Zampini #endif 130208122e43SStefano Zampini if (pcbddc->dbg_flag) { 13031b968477SStefano Zampini PetscInt maxneigs_r; 1304b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 13059b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 130608122e43SStefano Zampini } 130708122e43SStefano Zampini PetscFunctionReturn(0); 130808122e43SStefano Zampini } 1309b1b3d7a2SStefano Zampini 1310674ae819SStefano Zampini #undef __FUNCT__ 1311c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1312c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1313c8587f34SStefano Zampini { 1314c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13158629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1316c8587f34SStefano Zampini PetscErrorCode ierr; 1317c8587f34SStefano Zampini 1318c8587f34SStefano Zampini PetscFunctionBegin; 1319f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 13205e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1321c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1322c8587f34SStefano Zampini 1323684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 13240fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1325684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1326c8587f34SStefano Zampini 1327c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1328b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1329c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1330c8587f34SStefano Zampini } 1331c8587f34SStefano Zampini 13328629588bSStefano Zampini /* 13338629588bSStefano Zampini Setup local correction and local part of coarse basis. 13348629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13358629588bSStefano Zampini */ 133647f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13378629588bSStefano Zampini 13388629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13398629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13408629588bSStefano Zampini 13418629588bSStefano Zampini /* free */ 13428629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1343c8587f34SStefano Zampini PetscFunctionReturn(0); 1344c8587f34SStefano Zampini } 1345c8587f34SStefano Zampini 1346c8587f34SStefano Zampini #undef __FUNCT__ 1347674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1348674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1349674ae819SStefano Zampini { 1350674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1351674ae819SStefano Zampini PetscErrorCode ierr; 1352674ae819SStefano Zampini 1353674ae819SStefano Zampini PetscFunctionBegin; 1354674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1355674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 135630368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1357674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1358674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1359785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1360674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1361f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1362f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1363785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 136463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 136563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1366674ae819SStefano Zampini PetscFunctionReturn(0); 1367674ae819SStefano Zampini } 1368674ae819SStefano Zampini 1369674ae819SStefano Zampini #undef __FUNCT__ 1370674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1371674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1372674ae819SStefano Zampini { 1373674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13744f1b2e48SStefano Zampini PetscInt i; 1375674ae819SStefano Zampini PetscErrorCode ierr; 1376674ae819SStefano Zampini 1377674ae819SStefano Zampini PetscFunctionBegin; 1378b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1379674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 138016909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 13811dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1382674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1383674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13844f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13854f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13864f1b2e48SStefano Zampini } 13874f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1388b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1389674ae819SStefano Zampini PetscFunctionReturn(0); 1390674ae819SStefano Zampini } 1391674ae819SStefano Zampini 1392674ae819SStefano Zampini #undef __FUNCT__ 1393674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1394674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1395674ae819SStefano Zampini { 1396674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1397674ae819SStefano Zampini PetscErrorCode ierr; 1398674ae819SStefano Zampini 1399674ae819SStefano Zampini PetscFunctionBegin; 1400674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 140158da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1402ca92afb2SStefano Zampini PetscScalar *array; 140306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 140406656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 140558da7f69SStefano Zampini } 1406674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1407674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 140815aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 140915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1410674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1411674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1412674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 141306656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1414674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1415674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14168ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1417674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1418674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1419674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1420f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1421f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1422f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1423f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1424727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 14250e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1426f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 142770cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 142881d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14290369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14301dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 14314f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14328b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1433ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1434ca92afb2SStefano Zampini PetscInt i; 1435ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1436ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1437ca92afb2SStefano Zampini } 1438ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1439ca92afb2SStefano Zampini } 14404f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1441674ae819SStefano Zampini PetscFunctionReturn(0); 1442674ae819SStefano Zampini } 1443674ae819SStefano Zampini 1444674ae819SStefano Zampini #undef __FUNCT__ 1445f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1446f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14476bfb1811SStefano Zampini { 14486bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14496bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14506bfb1811SStefano Zampini VecType impVecType; 14514f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14526bfb1811SStefano Zampini PetscErrorCode ierr; 14536bfb1811SStefano Zampini 14546bfb1811SStefano Zampini PetscFunctionBegin; 14556c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1456e7b262bdSStefano Zampini /* get sizes */ 14574f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1458b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14596bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1460e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1461e7b262bdSStefano Zampini /* R nodes */ 1462e7b262bdSStefano Zampini old_size = -1; 1463e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1464e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1465e7b262bdSStefano Zampini } 1466e7b262bdSStefano Zampini if (n_R != old_size) { 1467e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1468e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14696bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14706bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14716bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14726bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1473e7b262bdSStefano Zampini } 1474e7b262bdSStefano Zampini /* local primal dofs */ 1475e7b262bdSStefano Zampini old_size = -1; 1476e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1477e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1478e7b262bdSStefano Zampini } 1479e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1480e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 148183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1482e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14836bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1484e7b262bdSStefano Zampini } 1485e7b262bdSStefano Zampini /* local explicit constraints */ 1486e7b262bdSStefano Zampini old_size = -1; 1487e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1488e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1489e7b262bdSStefano Zampini } 1490e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1491e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 149283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 149383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 149483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 149583b7ccabSStefano Zampini } 14966bfb1811SStefano Zampini PetscFunctionReturn(0); 14976bfb1811SStefano Zampini } 14986bfb1811SStefano Zampini 14996bfb1811SStefano Zampini #undef __FUNCT__ 150047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 150147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 150288ebb749SStefano Zampini { 150325084f0cSStefano Zampini PetscErrorCode ierr; 150425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 150588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 150688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1507d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 150825084f0cSStefano Zampini /* submatrices of local problem */ 150980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 151006656605SStefano Zampini /* submatrices of local coarse problem */ 151106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 151225084f0cSStefano Zampini /* working matrices */ 151306656605SStefano Zampini Mat C_CR; 151425084f0cSStefano Zampini /* additional working stuff */ 151506656605SStefano Zampini PC pc_R; 15164f1b2e48SStefano Zampini Mat F; 15175cbda25cSStefano Zampini Vec dummy_vec; 1518a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 151925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 152006656605SStefano Zampini PetscScalar *work; 152106656605SStefano Zampini PetscInt *idx_V_B; 1522ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 152306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1524ffd830a3SStefano Zampini 152525084f0cSStefano Zampini /* some shortcuts to scalars */ 152606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 152788ebb749SStefano Zampini 152888ebb749SStefano Zampini PetscFunctionBegin; 1529ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1530ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1531ffd830a3SStefano Zampini } 1532ffd830a3SStefano Zampini 1533ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1534b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15354f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1536b371cd4fSStefano Zampini n_B = pcis->n_B; 1537b371cd4fSStefano Zampini n_D = pcis->n - n_B; 153888ebb749SStefano Zampini n_R = pcis->n - n_vertices; 153988ebb749SStefano Zampini 154088ebb749SStefano Zampini /* vertices in boundary numbering */ 1541785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15420e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 15436c4ed002SBarry 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); 154488ebb749SStefano Zampini 154506656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1546019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 154706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 154806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 154906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 155106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 155206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 155306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 155406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 155506656605SStefano Zampini 155606656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 155706656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 155806656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 155906656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 156006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1561ffd830a3SStefano Zampini lda_rhs = n_R; 1562a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 156306656605SStefano Zampini if (isLU || isILU || isCHOL) { 156406656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1565df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1566df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1567d62866d3SStefano Zampini MatFactorType type; 1568d62866d3SStefano Zampini 1569df4d28bfSStefano Zampini F = reuse_solver->F; 15706816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1571d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1572ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 157322db5ddcSStefano Zampini need_benign_correction = (PetscBool)(!!reuse_solver->benign_n); 157406656605SStefano Zampini } else { 157506656605SStefano Zampini F = NULL; 157606656605SStefano Zampini } 157706656605SStefano Zampini 1578ffd830a3SStefano Zampini /* allocate workspace */ 1579ffd830a3SStefano Zampini n = 0; 1580ffd830a3SStefano Zampini if (n_constraints) { 1581ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1582ffd830a3SStefano Zampini } 1583ffd830a3SStefano Zampini if (n_vertices) { 1584ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1585ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1586ffd830a3SStefano Zampini } 1587ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1588ffd830a3SStefano Zampini 15895cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 15905cbda25cSStefano Zampini dummy_vec = NULL; 15915cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 15925cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 15935cbda25cSStefano Zampini } 15945cbda25cSStefano Zampini 159588ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 159688ebb749SStefano Zampini if (n_constraints) { 159772b8c272SStefano Zampini Mat M1,M2,M3,C_B; 159806656605SStefano Zampini IS is_aux; 159980677318SStefano Zampini PetscScalar *array,*array2; 160006656605SStefano Zampini 1601f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 160280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 160388ebb749SStefano Zampini 160425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 160525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 16068ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 160772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 160888ebb749SStefano Zampini 160980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 161080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1611ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 161288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 161306656605SStefano Zampini const PetscScalar *row_cmat_values; 161406656605SStefano Zampini const PetscInt *row_cmat_indices; 161506656605SStefano Zampini PetscInt size_of_constraint,j; 161688ebb749SStefano Zampini 161706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1619ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 162006656605SStefano Zampini } 162106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 162206656605SStefano Zampini } 1623ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 162406656605SStefano Zampini if (F) { 162506656605SStefano Zampini Mat B; 162606656605SStefano Zampini 1627ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1628a3df083aSStefano Zampini if (need_benign_correction) { 1629df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1630a3df083aSStefano Zampini 163172b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 163272b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1633a3df083aSStefano Zampini } 163480677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1635a3df083aSStefano Zampini if (need_benign_correction) { 1636a3df083aSStefano Zampini PetscScalar *marr; 1637df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1638a3df083aSStefano Zampini 1639a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 16405cbda25cSStefano Zampini if (lda_rhs != n_R) { 16415cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 16425cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 16435cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 16445cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 16455cbda25cSStefano Zampini } 16465cbda25cSStefano Zampini } else { 1647a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1648a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 16495cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1650a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1651a3df083aSStefano Zampini } 16525cbda25cSStefano Zampini } 1653a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1654a3df083aSStefano Zampini } 165506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 165606656605SStefano Zampini } else { 165780677318SStefano Zampini PetscScalar *marr; 165880677318SStefano Zampini 165980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 166006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1661ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1662ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 166306656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 166406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 166506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 166606656605SStefano Zampini } 166780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 166806656605SStefano Zampini } 166980677318SStefano Zampini if (!pcbddc->switch_static) { 167080677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 167180677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 167280677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 167380677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1674ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 167580677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 167680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167880677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 167980677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 168080677318SStefano Zampini } 168180677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 168280677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 168372b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 168480677318SStefano Zampini } else { 1685ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1686ffd830a3SStefano Zampini IS dummy; 1687ffd830a3SStefano Zampini 1688ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 168972b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1690ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1691ffd830a3SStefano Zampini } else { 169280677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 169380677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1694ffd830a3SStefano Zampini } 169525084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 169680677318SStefano Zampini } 169780677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 169880677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 169980677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 170006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 170106656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 170280677318SStefano Zampini if (isCHOL) { 170380677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 170480677318SStefano Zampini } else { 170525084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 170680677318SStefano Zampini } 170780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 170806656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 170925084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 171025084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 171125084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 171280677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 171372b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 171472b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 171506656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 171606656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1717f4ddd8eeSStefano Zampini } 1718fc227af8SStefano Zampini 1719fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 172088ebb749SStefano Zampini if (n_vertices) { 172106656605SStefano Zampini IS is_aux; 17223a50541eSStefano Zampini 1723df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 17246816873aSStefano Zampini IS tis; 17256816873aSStefano Zampini 17266816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17276816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17286816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17296816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17306816873aSStefano Zampini } else { 17313a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17326816873aSStefano Zampini } 17339577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17349577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 173504708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 173625084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 173788ebb749SStefano Zampini } 173888ebb749SStefano Zampini 173988ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1740f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 174106656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 174206656605SStefano Zampini if (pcbddc->coarse_phi_D) { 174306656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 174406656605SStefano Zampini } 1745f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 174606656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 174706656605SStefano Zampini PetscScalar *marray; 174806656605SStefano Zampini 174906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 175006656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1751f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1752f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1753f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1754f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1755f4ddd8eeSStefano Zampini } 1756f4ddd8eeSStefano Zampini } 175706656605SStefano Zampini 1758f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 175906656605SStefano Zampini PetscScalar *marray; 176088ebb749SStefano Zampini 176106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17628eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 176306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 176488ebb749SStefano Zampini } 17653301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 176606656605SStefano Zampini n *= 2; 176788ebb749SStefano Zampini } 176806656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 176906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 177006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17718eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 177206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 177306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 177488ebb749SStefano Zampini } 17753301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 177606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17778eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 177806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 177906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 178088ebb749SStefano Zampini } 178188ebb749SStefano Zampini } else { 1782c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1783c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17841b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1785c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1786c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1787c0553b1fSStefano Zampini } 178888ebb749SStefano Zampini } 178906656605SStefano Zampini } 1790019a44ceSStefano Zampini 179106656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17924f1b2e48SStefano Zampini p0_lidx_I = NULL; 17934f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1794d12edf2fSStefano Zampini const PetscInt *idxs; 1795d12edf2fSStefano Zampini 1796d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17974f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17984f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17994f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 18004f1b2e48SStefano Zampini } 1801d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1802d12edf2fSStefano Zampini } 1803d16cbb6bSStefano Zampini 180406656605SStefano Zampini /* vertices */ 180506656605SStefano Zampini if (n_vertices) { 180616f15bc4SStefano Zampini 1807af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 180804708bb6SStefano Zampini 180916f15bc4SStefano Zampini if (n_R) { 181014393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 181106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 181216f15bc4SStefano Zampini PetscScalar *x,*y; 181304708bb6SStefano Zampini PetscBool isseqaij; 181406656605SStefano Zampini 181521eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 181614393ed6SStefano Zampini if (need_benign_correction) { 181714393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 181814393ed6SStefano Zampini IS is_p0; 181914393ed6SStefano Zampini PetscInt *idxs_p0,n; 182014393ed6SStefano Zampini 182114393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 182214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 182314393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 1824af25d912SStefano 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); 182514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 182614393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 182714393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 182814393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 182914393ed6SStefano Zampini } 183014393ed6SStefano Zampini 1831ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1832af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1833ffd830a3SStefano Zampini } else { 1834ca92afb2SStefano Zampini PetscScalar *av,*array; 1835ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1836ca92afb2SStefano Zampini PetscInt n; 1837ca92afb2SStefano Zampini PetscBool flg_row; 1838ffd830a3SStefano Zampini 1839ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1840ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 18419d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1842ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1843ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1844ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1845ca92afb2SStefano Zampini PetscInt j; 1846ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1847ffd830a3SStefano Zampini } 1848ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1849ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1850ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1851ffd830a3SStefano Zampini } 1852ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1853a3df083aSStefano Zampini if (need_benign_correction) { 1854df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1855a3df083aSStefano Zampini PetscScalar *marr; 1856a3df083aSStefano Zampini 1857a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 185814393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 185914393ed6SStefano Zampini 186014393ed6SStefano Zampini | 0 0 0 | (V) 186114393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 186214393ed6SStefano Zampini | 0 0 -1 | (p0) 186314393ed6SStefano Zampini 186414393ed6SStefano Zampini */ 1865df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 186614393ed6SStefano Zampini const PetscScalar *vals; 186714393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 186814393ed6SStefano Zampini PetscInt n,j,nz; 186914393ed6SStefano Zampini 1870df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1871df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 187214393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 187314393ed6SStefano Zampini for (j=0;j<n;j++) { 187414393ed6SStefano Zampini PetscScalar val = vals[j]; 187514393ed6SStefano Zampini PetscInt k,col = idxs[j]; 187614393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 187714393ed6SStefano Zampini } 187814393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1879df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 188014393ed6SStefano Zampini } 188172b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 188272b8c272SStefano Zampini } 188372b8c272SStefano Zampini if (F) { 188414393ed6SStefano Zampini /* need to correct the rhs */ 188572b8c272SStefano Zampini if (need_benign_correction) { 188672b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 188772b8c272SStefano Zampini PetscScalar *marr; 188872b8c272SStefano Zampini 188972b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 18905cbda25cSStefano Zampini if (lda_rhs != n_R) { 18915cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18925cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18935cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 18945cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18955cbda25cSStefano Zampini } 18965cbda25cSStefano Zampini } else { 1897a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1898a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18995cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1900a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1901a3df083aSStefano Zampini } 19025cbda25cSStefano Zampini } 1903a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1904a3df083aSStefano Zampini } 190506656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 190614393ed6SStefano Zampini /* need to correct the solution */ 1907a3df083aSStefano Zampini if (need_benign_correction) { 1908df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1909a3df083aSStefano Zampini PetscScalar *marr; 1910a3df083aSStefano Zampini 1911a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 19125cbda25cSStefano Zampini if (lda_rhs != n_R) { 19135cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 19145cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19155cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19165cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19175cbda25cSStefano Zampini } 19185cbda25cSStefano Zampini } else { 1919a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1920a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19215cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1922a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1923a3df083aSStefano Zampini } 19245cbda25cSStefano Zampini } 1925a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1926a3df083aSStefano Zampini } 192706656605SStefano Zampini } else { 192806656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 192906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1930ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1931ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 193206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 193306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 193406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 193506656605SStefano Zampini } 193606656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 193706656605SStefano Zampini } 193880677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1939ffd830a3SStefano Zampini /* S_VV and S_CV */ 194006656605SStefano Zampini if (n_constraints) { 194106656605SStefano Zampini Mat B; 194280677318SStefano Zampini 1943ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 194480677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1945ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1946ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 194780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194980677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 195080677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 195180677318SStefano Zampini } 1952ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 195380677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 195480677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1955ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 195680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 195706656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1958ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1959ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 196006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 196106656605SStefano Zampini } 196204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 196304708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 1964511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 196504708bb6SStefano Zampini } 1966ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1967ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1968ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1969ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1970ffd830a3SStefano Zampini } 197106656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 197214393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 197314393ed6SStefano Zampini if (need_benign_correction) { 1974df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 197514393ed6SStefano Zampini PetscScalar *marr,*sums; 197614393ed6SStefano Zampini 197714393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 197814393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1979df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 198014393ed6SStefano Zampini const PetscScalar *vals; 198114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 198214393ed6SStefano Zampini PetscInt n,j,nz; 198314393ed6SStefano Zampini 1984df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1985df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 198614393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 198714393ed6SStefano Zampini PetscInt k; 198814393ed6SStefano Zampini sums[j] = 0.; 198914393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 199014393ed6SStefano Zampini } 199114393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 199214393ed6SStefano Zampini for (j=0;j<n;j++) { 199314393ed6SStefano Zampini PetscScalar val = vals[j]; 199414393ed6SStefano Zampini PetscInt k; 199514393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 199614393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 199714393ed6SStefano Zampini } 199814393ed6SStefano Zampini } 199914393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 2000df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 200114393ed6SStefano Zampini } 200214393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 200314393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 200414393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 200514393ed6SStefano Zampini } 200680677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 200706656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 200806656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 200906656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 201006656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 201106656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 201206656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 201306656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2014d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2015019a44ceSStefano Zampini } else { 2016d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2017d16cbb6bSStefano Zampini } 201821eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2019d16cbb6bSStefano Zampini 202006656605SStefano Zampini /* coarse basis functions */ 202106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 202216f15bc4SStefano Zampini PetscScalar *y; 202316f15bc4SStefano Zampini 2024ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 202506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 202706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 203006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 203106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 203206656605SStefano Zampini 203306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20344f1b2e48SStefano Zampini PetscInt j; 20354f1b2e48SStefano Zampini 203606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 203806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 204006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20414f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 204206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 204306656605SStefano Zampini } 204406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 204506656605SStefano Zampini } 204604708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 204704708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 204806656605SStefano Zampini } 20495cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 205006656605SStefano Zampini 205106656605SStefano Zampini if (n_constraints) { 205206656605SStefano Zampini Mat B; 205306656605SStefano Zampini 2054ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 205506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 205706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205806656605SStefano Zampini if (n_vertices) { 205980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 206080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 206180677318SStefano Zampini } else { 206280677318SStefano Zampini Mat S_VCt; 206380677318SStefano Zampini 2064ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2065ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206672b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2067ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2068ffd830a3SStefano Zampini } 206980677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 207080677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 207180677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 207280677318SStefano Zampini } 207306656605SStefano Zampini } 207406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 207506656605SStefano Zampini /* coarse basis functions */ 207606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 207706656605SStefano Zampini PetscScalar *y; 207806656605SStefano Zampini 2079ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 208006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 208106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 208206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 208506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 208606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20874f1b2e48SStefano Zampini PetscInt j; 20884f1b2e48SStefano Zampini 208906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 209006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 209106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 209206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 209306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20944f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 209506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 209606656605SStefano Zampini } 209706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209806656605SStefano Zampini } 209906656605SStefano Zampini } 210080677318SStefano Zampini if (n_constraints) { 210180677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 210280677318SStefano Zampini } 21034f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 210472b8c272SStefano Zampini 210572b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 210672b8c272SStefano Zampini if (pcbddc->benign_n) { 210772b8c272SStefano Zampini Mat B0_B,B0_BPHI; 210872b8c272SStefano Zampini IS is_dummy; 210972b8c272SStefano Zampini PetscScalar *data; 211072b8c272SStefano Zampini PetscInt j; 211172b8c272SStefano Zampini 211272b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 211372b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 211472b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 211572b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 211686c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 211772b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 211872b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 211972b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 212072b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 212172b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 212272b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 212372b8c272SStefano Zampini } 212472b8c272SStefano Zampini } 212572b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 212672b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 212772b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 212872b8c272SStefano Zampini } 2129019a44ceSStefano Zampini 213006656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 21313301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2132ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2133ffd830a3SStefano Zampini PetscScalar *marray; 213406656605SStefano Zampini 213506656605SStefano Zampini if (n_constraints) { 2136ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 213706656605SStefano Zampini 2138af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 213906656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2140ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 214116f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 214206656605SStefano Zampini if (n_vertices) { 2143ffd830a3SStefano Zampini Mat S_VCT; 214406656605SStefano Zampini 214506656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2146ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 214716f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 214806656605SStefano Zampini } 2149ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 215006656605SStefano Zampini } 215116f15bc4SStefano Zampini if (n_vertices && n_R) { 2152ffd830a3SStefano Zampini PetscScalar *av,*marray; 2153ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2154ffd830a3SStefano Zampini PetscInt n; 2155ffd830a3SStefano Zampini PetscBool flg_row; 215606656605SStefano Zampini 2157ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2158af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2159ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2160ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2161ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2162ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2163ffd830a3SStefano Zampini PetscInt j; 2164ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2165ffd830a3SStefano Zampini } 2166ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2167ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2168ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 216906656605SStefano Zampini } 217006656605SStefano Zampini 2171ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2172ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2173ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2174ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2175ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 217606656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 217706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 217906656605SStefano Zampini } 2180ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2181ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2182ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2183ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2184ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2185ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2186ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2187ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 218806656605SStefano Zampini } 2189ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 219006656605SStefano Zampini /* coarse basis functions */ 219106656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 219206656605SStefano Zampini PetscScalar *y; 219306656605SStefano Zampini 2194ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 219506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 219706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219906656605SStefano Zampini if (i<n_vertices) { 220006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 220106656605SStefano Zampini } 220206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 220306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 220406656605SStefano Zampini 220506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 220606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 220806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 221006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 221106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 221206656605SStefano Zampini } 221306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 221406656605SStefano Zampini } 2215ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2216ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 221706656605SStefano Zampini } 2218d62866d3SStefano Zampini /* free memory */ 221988ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 222006656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 222106656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 222206656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 222306656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2224d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2225d62866d3SStefano Zampini if (n_vertices) { 2226d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2227d62866d3SStefano Zampini } 2228d62866d3SStefano Zampini if (n_constraints) { 2229d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2230d62866d3SStefano Zampini } 223188ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 223288ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 223388ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2234d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 223588ebb749SStefano Zampini Mat coarse_sub_mat; 223625084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 223788ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 223888ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 223988ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22408bec7fa6SStefano Zampini Mat C_B,CPHI; 22418bec7fa6SStefano Zampini IS is_dummy; 22428bec7fa6SStefano Zampini Vec mones; 224388ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 224488ebb749SStefano Zampini PetscReal real_value; 224588ebb749SStefano Zampini 2246a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2247a3df083aSStefano Zampini Mat A; 2248a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2249a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2250a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2251a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2252a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2253a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2254a3df083aSStefano Zampini } else { 225588ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 225688ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 225788ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 225888ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2259a3df083aSStefano Zampini } 226088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 226188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2262ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 226488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 226588ebb749SStefano Zampini } 226688ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 226788ebb749SStefano Zampini 226825084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22693301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 227025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2271ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 227288ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 227788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227888ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 228088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228188ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 228388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228488ebb749SStefano Zampini } else { 228588ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 228688ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 228788ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 228988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 229088ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 229188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 229288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 229388ebb749SStefano Zampini } 229488ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229588ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2297511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 22984f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2299fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2300d12edf2fSStefano Zampini PetscScalar *data,*data2; 23014f1b2e48SStefano Zampini PetscInt j; 2302d12edf2fSStefano Zampini 23034f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2304fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2305d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 230686c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2307d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2308d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 23094f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 23104f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2311d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 23124f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 23134f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 23144f1b2e48SStefano Zampini } 2315d12edf2fSStefano Zampini } 2316d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2317d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2318d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2319d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2320d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2321d12edf2fSStefano Zampini } 2322d12edf2fSStefano Zampini #if 0 2323d12edf2fSStefano Zampini { 2324d12edf2fSStefano Zampini PetscViewer viewer; 2325d12edf2fSStefano Zampini char filename[256]; 2326ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2327d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2328d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2329ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2330ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2331ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2332d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 233372b8c272SStefano Zampini if (save_change) { 233472b8c272SStefano Zampini Mat phi_B; 233572b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 233672b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 233772b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 233872b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 233972b8c272SStefano Zampini } else { 2340ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2341ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 234272b8c272SStefano Zampini } 2343ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2344ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2345ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2346ffd830a3SStefano Zampini } 2347ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2348ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2349ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2350ffd830a3SStefano Zampini } 235172b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2352ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2353ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2354ffd830a3SStefano Zampini } 2355d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2356d12edf2fSStefano Zampini } 2357d12edf2fSStefano Zampini #endif 235881d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23598bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23601575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 236106656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23628bec7fa6SStefano Zampini 23638bec7fa6SStefano Zampini /* check constraints */ 2364a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2365a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 23664f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23678bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2368a00504b5SStefano Zampini } else { 2369a00504b5SStefano Zampini PetscScalar *data; 2370a00504b5SStefano Zampini Mat tmat; 2371a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2372a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2373a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2374a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2375a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2376a00504b5SStefano Zampini } 23778bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23788bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23798bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23808bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2381bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2382ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2383bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2384bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2385bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2386bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2387bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 238888ebb749SStefano Zampini } 23898bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23908bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23918bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23928bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 239325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 239488ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 239588ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 239688ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 239788ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 239888ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 239988ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 240088ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 240188ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 240288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 240388ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2404ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 240588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 240688ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 240788ebb749SStefano Zampini } 240888ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 240988ebb749SStefano Zampini } 24108629588bSStefano Zampini /* get back data */ 24118629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 241288ebb749SStefano Zampini PetscFunctionReturn(0); 241388ebb749SStefano Zampini } 241488ebb749SStefano Zampini 241588ebb749SStefano Zampini #undef __FUNCT__ 2416d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2417d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2418aa0d41d4SStefano Zampini { 2419d65f70fdSStefano Zampini Mat *work_mat; 2420d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2421d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2422d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2423aa0d41d4SStefano Zampini PetscErrorCode ierr; 2424aa0d41d4SStefano Zampini 2425aa0d41d4SStefano Zampini PetscFunctionBegin; 2426d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2427d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2428d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2429d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2430aa0d41d4SStefano Zampini 2431d65f70fdSStefano Zampini if (!rsorted) { 2432906d46d4SStefano Zampini const PetscInt *idxs; 2433906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2434aa0d41d4SStefano Zampini 2435d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2436d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2437d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2438d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2439aa0d41d4SStefano Zampini } 2440d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2441d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2442d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2443d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2444aa0d41d4SStefano Zampini } 2445d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2446d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2447d65f70fdSStefano Zampini } else { 2448d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2449d65f70fdSStefano Zampini isrow_s = isrow; 2450aa0d41d4SStefano Zampini } 2451906d46d4SStefano Zampini 2452d65f70fdSStefano Zampini if (!csorted) { 2453d65f70fdSStefano Zampini if (isrow == iscol) { 2454d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2455d65f70fdSStefano Zampini iscol_s = isrow_s; 2456d65f70fdSStefano Zampini } else { 2457d65f70fdSStefano Zampini const PetscInt *idxs; 2458d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2459906d46d4SStefano Zampini 2460d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2461d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2462d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2463d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2464d65f70fdSStefano Zampini } 2465d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2466d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2467d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2468d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2469d65f70fdSStefano Zampini } 2470d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2471d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2472d65f70fdSStefano Zampini } 2473d65f70fdSStefano Zampini } else { 2474d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2475d65f70fdSStefano Zampini iscol_s = iscol; 2476d65f70fdSStefano Zampini } 2477d65f70fdSStefano Zampini 2478d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2479d65f70fdSStefano Zampini 2480d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2481906d46d4SStefano Zampini Mat new_mat; 2482d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2483906d46d4SStefano Zampini 2484d65f70fdSStefano Zampini if (!rsorted) { 2485d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2486d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2487d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2488d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2489906d46d4SStefano Zampini } 2490d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2491d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2492d65f70fdSStefano Zampini } else { 2493d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2494906d46d4SStefano Zampini } 2495d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2496d65f70fdSStefano Zampini 2497d65f70fdSStefano Zampini if (!csorted) { 2498d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2499d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2500d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2501d65f70fdSStefano Zampini } else { 2502d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2503d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2504d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2505d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2506d65f70fdSStefano Zampini } 2507d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2508d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2509d65f70fdSStefano Zampini } 2510d65f70fdSStefano Zampini } else { 2511d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2512d65f70fdSStefano Zampini } 2513d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2514d65f70fdSStefano Zampini 2515d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2516d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2517d65f70fdSStefano Zampini work_mat[0] = new_mat; 2518d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2519d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2520d65f70fdSStefano Zampini } 2521d65f70fdSStefano Zampini 2522d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2523d65f70fdSStefano Zampini *B = work_mat[0]; 2524d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2525d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2526d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2527d65f70fdSStefano Zampini PetscFunctionReturn(0); 2528d65f70fdSStefano Zampini } 2529d65f70fdSStefano Zampini 2530d65f70fdSStefano Zampini #undef __FUNCT__ 25315e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 25325e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2533aa0d41d4SStefano Zampini { 2534aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 25355e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2536d65f70fdSStefano Zampini Mat new_mat; 25375e8657edSStefano Zampini IS is_local,is_global; 2538d65f70fdSStefano Zampini PetscInt local_size; 2539d65f70fdSStefano Zampini PetscBool isseqaij; 2540aa0d41d4SStefano Zampini PetscErrorCode ierr; 2541aa0d41d4SStefano Zampini 2542aa0d41d4SStefano Zampini PetscFunctionBegin; 2543aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25445e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25455e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2546b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2547aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2548d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2549aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2550906d46d4SStefano Zampini 2551906d46d4SStefano Zampini /* check */ 2552906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2553906d46d4SStefano Zampini Vec x,x_change; 2554906d46d4SStefano Zampini PetscReal error; 2555906d46d4SStefano Zampini 25565e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2557906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25585e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2559e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2560e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2561d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2562e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2563e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2564906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2565906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2566906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2567906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2568906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2569906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2570906d46d4SStefano Zampini } 2571906d46d4SStefano Zampini 257222d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25739b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 257422d5777bSStefano Zampini if (isseqaij) { 2575a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2576a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2577aa0d41d4SStefano Zampini } else { 2578a00504b5SStefano Zampini Mat work_mat; 25791cf9b237SStefano Zampini 2580a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2581aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2582a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2583aa0d41d4SStefano Zampini } 25843301b35fSStefano Zampini if (matis->A->symmetric_set) { 25853301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2586e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25873301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2588e496cd5dSStefano Zampini #endif 25893301b35fSStefano Zampini } 2590d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2591aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2592aa0d41d4SStefano Zampini } 2593aa0d41d4SStefano Zampini 2594aa0d41d4SStefano Zampini #undef __FUNCT__ 2595a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25968ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2597a64d13efSStefano Zampini { 2598a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2599a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2600d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 260153892102SStefano Zampini PetscInt *idx_R_local=NULL; 26023a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 26033a50541eSStefano Zampini PetscInt vbs,bs; 26046816873aSStefano Zampini PetscBT bitmask=NULL; 2605a64d13efSStefano Zampini PetscErrorCode ierr; 2606a64d13efSStefano Zampini 2607a64d13efSStefano Zampini PetscFunctionBegin; 2608b23d619eSStefano Zampini /* 2609b23d619eSStefano Zampini No need to setup local scatters if 2610b23d619eSStefano Zampini - primal space is unchanged 2611b23d619eSStefano Zampini AND 2612b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2613b23d619eSStefano Zampini AND 2614b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2615b23d619eSStefano Zampini */ 2616b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2617f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2618f4ddd8eeSStefano Zampini } 2619f4ddd8eeSStefano Zampini /* destroy old objects */ 2620f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2621f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2622f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2623a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2624b371cd4fSStefano Zampini n_B = pcis->n_B; 2625b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2626b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 26273a50541eSStefano Zampini 2628a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 26296816873aSStefano Zampini 263053892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2631df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2632854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2633a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2634a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 26350e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2636a64d13efSStefano Zampini } 2637a64d13efSStefano Zampini 2638a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26394641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26406816873aSStefano Zampini idx_R_local[n_R++] = i; 2641a64d13efSStefano Zampini } 2642a64d13efSStefano Zampini } 2643df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2644df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26456816873aSStefano Zampini 2646df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2647df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26486816873aSStefano Zampini } 26493a50541eSStefano Zampini 26503a50541eSStefano Zampini /* Block code */ 26513a50541eSStefano Zampini vbs = 1; 26523a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26533a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26543a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26553a50541eSStefano Zampini PetscInt *vary; 2656df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2657785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26583a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2659d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2660d3df7717SStefano 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 */ 26610e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2662d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26633a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26643a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26653a50541eSStefano Zampini break; 26663a50541eSStefano Zampini } 26673a50541eSStefano Zampini } 2668d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2669d3df7717SStefano Zampini } else { 2670d3df7717SStefano Zampini /* Verify directly the R set */ 2671d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2672d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2673d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2674d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2675d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2676d3df7717SStefano Zampini break; 2677d3df7717SStefano Zampini } 2678d3df7717SStefano Zampini } 2679d3df7717SStefano Zampini } 2680d3df7717SStefano Zampini } 26813a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26823a50541eSStefano Zampini vbs = bs; 26833a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26843a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26853a50541eSStefano Zampini } 26863a50541eSStefano Zampini } 26873a50541eSStefano Zampini } 26883a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2689df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2690df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 269153892102SStefano Zampini 2692df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2693df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 269453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2695df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 269653892102SStefano Zampini } else { 26973a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 269853892102SStefano Zampini } 2699a64d13efSStefano Zampini 2700a64d13efSStefano Zampini /* print some info if requested */ 2701a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2702a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2703a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 27041575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2705a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2706a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 27074f1b2e48SStefano 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); 2708a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2709a64d13efSStefano Zampini } 2710a64d13efSStefano Zampini 2711a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2712df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 27136816873aSStefano Zampini IS is_aux1,is_aux2; 27146816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 27156816873aSStefano Zampini 27163a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2717854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2718854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2719a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 27204641a718SStefano Zampini for (i=0; i<n_D; i++) { 27214641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 27224641a718SStefano Zampini } 2723a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2724a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27254641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 27264641a718SStefano Zampini aux_array1[j++] = i; 2727a64d13efSStefano Zampini } 2728a64d13efSStefano Zampini } 2729a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2730a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2731a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 27324641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 27334641a718SStefano Zampini aux_array2[j++] = i; 2734a64d13efSStefano Zampini } 2735a64d13efSStefano Zampini } 2736a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2737a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2738a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2739a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2740a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2741a64d13efSStefano Zampini 27428eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2743785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2744a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27454641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27464641a718SStefano Zampini aux_array1[j++] = i; 2747a64d13efSStefano Zampini } 2748a64d13efSStefano Zampini } 2749a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2750a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2751a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2752a64d13efSStefano Zampini } 27534641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27543a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2755d62866d3SStefano Zampini } else { 2756df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27576816873aSStefano Zampini IS tis; 27586816873aSStefano Zampini PetscInt schur_size; 27596816873aSStefano Zampini 2760df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27616816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2762df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27636816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27646816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27656816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27666816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27676816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2768d62866d3SStefano Zampini } 2769d62866d3SStefano Zampini } 2770a64d13efSStefano Zampini PetscFunctionReturn(0); 2771a64d13efSStefano Zampini } 2772a64d13efSStefano Zampini 2773304d26faSStefano Zampini 2774304d26faSStefano Zampini #undef __FUNCT__ 2775304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2776684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2777304d26faSStefano Zampini { 2778304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2779304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2780304d26faSStefano Zampini PC pc_temp; 2781304d26faSStefano Zampini Mat A_RR; 2782f4ddd8eeSStefano Zampini MatReuse reuse; 2783304d26faSStefano Zampini PetscScalar m_one = -1.0; 2784304d26faSStefano Zampini PetscReal value; 278504708bb6SStefano Zampini PetscInt n_D,n_R; 27869577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2787304d26faSStefano Zampini PetscErrorCode ierr; 2788e604994aSStefano Zampini /* prefixes stuff */ 2789312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2790e604994aSStefano Zampini size_t len; 2791304d26faSStefano Zampini 2792304d26faSStefano Zampini PetscFunctionBegin; 2793304d26faSStefano Zampini 2794e604994aSStefano Zampini /* compute prefixes */ 2795e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2796e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2797e604994aSStefano Zampini if (!pcbddc->current_level) { 2798e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2799e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2800e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2801e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2802e604994aSStefano Zampini } else { 2803e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2804312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2805e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2806e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2807312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2808312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 280934d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 281034d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2811e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2812e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2813e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2814e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2815e604994aSStefano Zampini } 2816e604994aSStefano Zampini 2817304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2818684f6988SStefano Zampini if (dirichlet) { 2819d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2820450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2821450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2822450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2823450f8f5eSStefano Zampini } 2824450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2825a3df083aSStefano Zampini Mat A_IIn; 2826a3df083aSStefano Zampini 2827a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2828a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2829a3df083aSStefano Zampini pcis->A_II = A_IIn; 2830a3df083aSStefano Zampini } 2831450f8f5eSStefano Zampini } 28323301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28333301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2834964fefecSStefano Zampini } 2835ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2836964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2837304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2838304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2839304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2840304d26faSStefano Zampini /* default */ 2841304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2842e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28439577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2844304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28459577ea80SStefano Zampini if (issbaij) { 28469577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28479577ea80SStefano Zampini } else { 2848304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28499577ea80SStefano Zampini } 2850304d26faSStefano Zampini /* Allow user's customization */ 2851304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2852304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2853304d26faSStefano Zampini } 2854d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2855df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2856df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2857d62866d3SStefano Zampini 2858df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2859d5574798SStefano Zampini } 2860304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2861304d26faSStefano Zampini if (!n_D) { 2862304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2863304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2864304d26faSStefano Zampini } 2865304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2866304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2867304d26faSStefano Zampini /* set ksp_D into pcis data */ 2868304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2869304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2870304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2871684f6988SStefano Zampini } 2872304d26faSStefano Zampini 2873304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2874684f6988SStefano Zampini A_RR = 0; 2875684f6988SStefano Zampini if (neumann) { 2876d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 287704708bb6SStefano Zampini PetscInt ibs,mbs; 287804708bb6SStefano Zampini PetscBool issbaij; 287904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2880f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28818ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2882f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2883f4ddd8eeSStefano Zampini PetscInt nn_R; 288481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2885f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2886f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2887f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2888f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2889f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2890f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2891f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2892727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2893f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2894f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2895f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2896f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2897f4ddd8eeSStefano Zampini } 2898f4ddd8eeSStefano Zampini } 2899f4ddd8eeSStefano Zampini /* last check */ 2900d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2901f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2902f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2903f4ddd8eeSStefano Zampini } 2904f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2905f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2906f4ddd8eeSStefano Zampini } 2907a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 2908af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2909af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 291004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 291104708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 291204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2915af732b37SStefano Zampini } else { 2916511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 29176816873aSStefano Zampini } 291804708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 291904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 292004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 292104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 292204708bb6SStefano Zampini } else { 2923511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 292404708bb6SStefano Zampini } 292504708bb6SStefano Zampini } 2926a00504b5SStefano Zampini /* extract A_RR */ 2927a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 2928a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2929a00504b5SStefano Zampini 2930a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 293116e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2932a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 293316e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 293416e386b8SStefano Zampini } else { 2935a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 2936a00504b5SStefano Zampini } 2937a00504b5SStefano Zampini } else { 2938a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2939a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 2940a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2941a00504b5SStefano Zampini } 2942a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 2943f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 294416e386b8SStefano Zampini } 29453301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29463301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29476816873aSStefano Zampini } 2948f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2949304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2950304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2951304d26faSStefano Zampini /* default */ 2952304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2953e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2954304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29559577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29569577ea80SStefano Zampini if (issbaij) { 29579577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29589577ea80SStefano Zampini } else { 2959304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29609577ea80SStefano Zampini } 2961304d26faSStefano Zampini /* Allow user's customization */ 2962304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2963304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2964304d26faSStefano Zampini } 2965304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2966304d26faSStefano Zampini if (!n_R) { 2967304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2968304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2969304d26faSStefano Zampini } 29705cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2971df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2972df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2973df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2974d62866d3SStefano Zampini 2975df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2976d62866d3SStefano Zampini } 2977304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2978304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2979684f6988SStefano Zampini } 2980304d26faSStefano Zampini 2981304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29820fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2983684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2984684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29851575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2986684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2987684f6988SStefano Zampini } 2988684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29890fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29900fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29910fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29920fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29930fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2994304d26faSStefano Zampini /* need to be adapted? */ 2995b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2996b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2997b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2998304d26faSStefano Zampini /* print info */ 2999304d26faSStefano Zampini if (pcbddc->dbg_flag) { 3000e604994aSStefano 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); 3001304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3002304d26faSStefano Zampini } 3003b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 3004298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 3005304d26faSStefano Zampini } 3006684f6988SStefano Zampini } 3007684f6988SStefano Zampini if (neumann) { /* Neumann */ 30080fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 30090fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 30100fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 30110fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 30120fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3013304d26faSStefano Zampini /* need to be adapted? */ 3014b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 3015b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3016304d26faSStefano Zampini /* print info */ 3017304d26faSStefano Zampini if (pcbddc->dbg_flag) { 3018e604994aSStefano 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); 3019304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3020304d26faSStefano Zampini } 3021b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 3022298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 3023304d26faSStefano Zampini } 30240fccc4e9SStefano Zampini } 3025684f6988SStefano Zampini } 30265cbda25cSStefano Zampini /* free Neumann problem's matrix */ 30275cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3028304d26faSStefano Zampini PetscFunctionReturn(0); 3029304d26faSStefano Zampini } 3030304d26faSStefano Zampini 3031304d26faSStefano Zampini #undef __FUNCT__ 3032ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 303380677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3034674ae819SStefano Zampini { 3035674ae819SStefano Zampini PetscErrorCode ierr; 3036674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3037be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3038674ae819SStefano Zampini 3039674ae819SStefano Zampini PetscFunctionBegin; 3040df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 304180677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 304220c7b377SStefano Zampini } 304380677318SStefano Zampini if (!pcbddc->switch_static) { 304480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 304580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 304680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 304720c7b377SStefano Zampini } 3048df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 304980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305120c7b377SStefano Zampini } else { 3052df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3053be83ff47SStefano Zampini 3054df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3055df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305620c7b377SStefano Zampini } 3057be83ff47SStefano Zampini } else { 305880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 306380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 306480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 306580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3067674ae819SStefano Zampini } 3068674ae819SStefano Zampini } 306972b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 307080677318SStefano Zampini if (applytranspose) { 307180677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 307280677318SStefano Zampini } else { 307380677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 307480677318SStefano Zampini } 3075be83ff47SStefano Zampini } else { 3076df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3077be83ff47SStefano Zampini 3078be83ff47SStefano Zampini if (applytranspose) { 3079df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3080be83ff47SStefano Zampini } else { 3081df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3082be83ff47SStefano Zampini } 3083be83ff47SStefano Zampini } 308480677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 308580677318SStefano Zampini if (!pcbddc->switch_static) { 3086df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 308780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3089be83ff47SStefano Zampini } else { 3090df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3091be83ff47SStefano Zampini 3092df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3093df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3094be83ff47SStefano Zampini } 309580677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 309680677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 309780677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 309880677318SStefano Zampini } 309980677318SStefano Zampini } else { 310080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 310580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 310680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 310780677318SStefano Zampini } 310880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 311080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 311180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3112674ae819SStefano Zampini } 3113674ae819SStefano Zampini PetscFunctionReturn(0); 3114674ae819SStefano Zampini } 3115674ae819SStefano Zampini 3116dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3117674ae819SStefano Zampini #undef __FUNCT__ 3118674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3119dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3120674ae819SStefano Zampini { 3121674ae819SStefano Zampini PetscErrorCode ierr; 3122674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3123674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3124674ae819SStefano Zampini const PetscScalar zero = 0.0; 3125674ae819SStefano Zampini 3126674ae819SStefano Zampini PetscFunctionBegin; 3127dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 31284fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3129dc359a40SStefano Zampini if (applytranspose) { 3130674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 31318eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3132dc359a40SStefano Zampini } else { 3133674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3134674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 313515aaf578SStefano Zampini } 31364fee134fSStefano Zampini } else { 31374fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 31384fee134fSStefano Zampini } 3139efc2fbd9SStefano Zampini 3140efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31414f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3142efc2fbd9SStefano Zampini PetscScalar *array; 31434f1b2e48SStefano Zampini PetscInt j; 3144efc2fbd9SStefano Zampini 3145efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31464f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3147efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3148efc2fbd9SStefano Zampini } 3149efc2fbd9SStefano Zampini 315012edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 315112edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 315212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315312edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315412edc857SStefano Zampini 31559f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 315612edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 315712edc857SStefano Zampini if (pcbddc->coarse_ksp) { 315851694757SStefano Zampini Mat coarse_mat; 3159964fefecSStefano Zampini Vec rhs,sol; 316051694757SStefano Zampini MatNullSpace nullsp; 316127b6a85dSStefano Zampini PetscBool isbddc = PETSC_FALSE; 3162964fefecSStefano Zampini 316327b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 316427b6a85dSStefano Zampini PC coarse_pc; 316527b6a85dSStefano Zampini 316627b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 316727b6a85dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);CHKERRQ(ierr); 316827b6a85dSStefano Zampini /* we need to propagate to coarser levels the need for a possible benign correction */ 316927b6a85dSStefano Zampini if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) { 317027b6a85dSStefano Zampini PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 317127b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_FALSE; 3172*3bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE; 317327b6a85dSStefano Zampini } 317427b6a85dSStefano Zampini } 3175964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3176964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 317751694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 317851694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 317951694757SStefano Zampini if (nullsp) { 318051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 318151694757SStefano Zampini } 318212edc857SStefano Zampini if (applytranspose) { 31832701bc32SStefano Zampini if (pcbddc->benign_apply_coarse_only) { /* need just to apply the coarse preconditioner */ 31842701bc32SStefano Zampini PC coarse_pc; 31852701bc32SStefano Zampini 31862701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 31872701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 31882701bc32SStefano Zampini ierr = PCApplyTranspose(coarse_pc,rhs,sol);CHKERRQ(ierr); 31892701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 31902701bc32SStefano Zampini } else { 3191964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 31922701bc32SStefano Zampini } 31932701bc32SStefano Zampini } else { 31942701bc32SStefano Zampini if (pcbddc->benign_apply_coarse_only) { /* need just to apply the coarse preconditioner */ 31952701bc32SStefano Zampini PC coarse_pc; 31962701bc32SStefano Zampini 31972701bc32SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 31982701bc32SStefano Zampini ierr = PCPreSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 31992701bc32SStefano Zampini ierr = PCApply(coarse_pc,rhs,sol);CHKERRQ(ierr); 32002701bc32SStefano Zampini ierr = PCPostSolve(coarse_pc,pcbddc->coarse_ksp);CHKERRQ(ierr); 320112edc857SStefano Zampini } else { 3202964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 320312edc857SStefano Zampini } 32042701bc32SStefano Zampini } 320527b6a85dSStefano Zampini /* we don't the benign correction at coarser levels anymore */ 320627b6a85dSStefano Zampini if (pcbddc->benign_have_null && isbddc) { 320727b6a85dSStefano Zampini PC coarse_pc; 320827b6a85dSStefano Zampini PC_BDDC* coarsepcbddc; 320927b6a85dSStefano Zampini 321027b6a85dSStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);CHKERRQ(ierr); 321127b6a85dSStefano Zampini coarsepcbddc = (PC_BDDC*)(coarse_pc->data); 321227b6a85dSStefano Zampini coarsepcbddc->benign_skip_correction = PETSC_TRUE; 3213*3bca92a6SStefano Zampini coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE; 321427b6a85dSStefano Zampini } 321551694757SStefano Zampini if (nullsp) { 321651694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 321751694757SStefano Zampini } 321812edc857SStefano Zampini } 3219674ae819SStefano Zampini 3220674ae819SStefano Zampini /* Local solution on R nodes */ 32214fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 322280677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 32239f00e9b4SStefano Zampini } 32249f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 32259f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 322612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3227674ae819SStefano Zampini 32284fee134fSStefano Zampini /* Sum contributions from the two levels */ 32294fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3230dc359a40SStefano Zampini if (applytranspose) { 3231dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3232dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3233dc359a40SStefano Zampini } else { 3234674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 32358eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3236dc359a40SStefano Zampini } 3237efc2fbd9SStefano Zampini /* store p0 */ 32384f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3239efc2fbd9SStefano Zampini PetscScalar *array; 32404f1b2e48SStefano Zampini PetscInt j; 3241efc2fbd9SStefano Zampini 3242efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 32434f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3244efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3245efc2fbd9SStefano Zampini } 32464fee134fSStefano Zampini } else { /* expand the coarse solution */ 32474fee134fSStefano Zampini if (applytranspose) { 32484fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32494fee134fSStefano Zampini } else { 32504fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32514fee134fSStefano Zampini } 32524fee134fSStefano Zampini } 3253674ae819SStefano Zampini PetscFunctionReturn(0); 3254674ae819SStefano Zampini } 3255674ae819SStefano Zampini 3256674ae819SStefano Zampini #undef __FUNCT__ 3257674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 325812edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3259674ae819SStefano Zampini { 3260674ae819SStefano Zampini PetscErrorCode ierr; 3261674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 326258da7f69SStefano Zampini PetscScalar *array; 326312edc857SStefano Zampini Vec from,to; 3264674ae819SStefano Zampini 3265674ae819SStefano Zampini PetscFunctionBegin; 326612edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 326712edc857SStefano Zampini from = pcbddc->coarse_vec; 326812edc857SStefano Zampini to = pcbddc->vec1_P; 326912edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 327012edc857SStefano Zampini Vec tvec; 327158da7f69SStefano Zampini 327258da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 327358da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 327412edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 327558da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 327658da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 327758da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 327812edc857SStefano Zampini } 327912edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 328012edc857SStefano Zampini from = pcbddc->vec1_P; 328112edc857SStefano Zampini to = pcbddc->coarse_vec; 328212edc857SStefano Zampini } 328312edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3284674ae819SStefano Zampini PetscFunctionReturn(0); 3285674ae819SStefano Zampini } 3286674ae819SStefano Zampini 3287674ae819SStefano Zampini #undef __FUNCT__ 3288674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 328912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3290674ae819SStefano Zampini { 3291674ae819SStefano Zampini PetscErrorCode ierr; 3292674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 329358da7f69SStefano Zampini PetscScalar *array; 329412edc857SStefano Zampini Vec from,to; 3295674ae819SStefano Zampini 3296674ae819SStefano Zampini PetscFunctionBegin; 329712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 329812edc857SStefano Zampini from = pcbddc->coarse_vec; 329912edc857SStefano Zampini to = pcbddc->vec1_P; 330012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 330112edc857SStefano Zampini from = pcbddc->vec1_P; 330212edc857SStefano Zampini to = pcbddc->coarse_vec; 330312edc857SStefano Zampini } 330412edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 330512edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 330612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 330712edc857SStefano Zampini Vec tvec; 330858da7f69SStefano Zampini 330912edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 331058da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 331158da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 331258da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 331358da7f69SStefano Zampini } 331458da7f69SStefano Zampini } else { 331558da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 331658da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 331712edc857SStefano Zampini } 331812edc857SStefano Zampini } 3319674ae819SStefano Zampini PetscFunctionReturn(0); 3320674ae819SStefano Zampini } 3321674ae819SStefano Zampini 3322984c4197SStefano Zampini /* uncomment for testing purposes */ 3323984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3324674ae819SStefano Zampini #undef __FUNCT__ 3325674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3326674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3327674ae819SStefano Zampini { 3328674ae819SStefano Zampini PetscErrorCode ierr; 3329674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3330674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3331674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3332984c4197SStefano Zampini /* one and zero */ 3333984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3334984c4197SStefano Zampini /* space to store constraints and their local indices */ 33359162d606SStefano Zampini PetscScalar *constraints_data; 33369162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 33379162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 33389162d606SStefano Zampini PetscInt *constraints_n; 3339984c4197SStefano Zampini /* iterators */ 3340b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3341984c4197SStefano Zampini /* BLAS integers */ 3342e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3343e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3344c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3345727cdba6SStefano Zampini /* reuse */ 33460e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 33470e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3348984c4197SStefano Zampini /* change of basis */ 3349b3d85658SStefano Zampini PetscBool qr_needed; 33509162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3351984c4197SStefano Zampini /* auxiliary stuff */ 335264efe560SStefano Zampini PetscInt *nnz,*is_indices; 33538a0068c3SStefano Zampini PetscInt ncc; 3354984c4197SStefano Zampini /* some quantities */ 335545a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3356a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3357984c4197SStefano Zampini 3358674ae819SStefano Zampini PetscFunctionBegin; 33598e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 33608e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 33618e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 336216909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3363088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3364088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 33650e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 33660e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 33670e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33680e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33690e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3370088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3371cf5a6209SStefano Zampini 3372cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33739162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3374cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3375cf5a6209SStefano Zampini const Vec *nearnullvecs; 3376cf5a6209SStefano Zampini Vec *localnearnullsp; 3377cf5a6209SStefano Zampini PetscScalar *array; 3378cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3379cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3380674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3381b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3382674ae819SStefano Zampini PetscScalar *work; 3383674ae819SStefano Zampini PetscReal *singular_vals; 3384674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3385674ae819SStefano Zampini PetscReal *rwork; 3386674ae819SStefano Zampini #endif 3387674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3388674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3389674ae819SStefano Zampini #else 3390964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3391964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3392674ae819SStefano Zampini #endif 3393674ae819SStefano Zampini 3394674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3395d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3396e4d548c7SStefano Zampini /* print some info */ 3397e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 3398e4d548c7SStefano Zampini PetscInt nv; 3399e4d548c7SStefano Zampini 3400e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 3401e4d548c7SStefano Zampini ierr = ISGetSize(ISForVertices,&nv);CHKERRQ(ierr); 3402e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3403e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3404e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3405e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);CHKERRQ(ierr); 3406e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);CHKERRQ(ierr); 3407e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3408e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3409e4d548c7SStefano Zampini } 3410e4d548c7SStefano Zampini 3411d06fc5fdSStefano Zampini /* free unneeded index sets */ 3412d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3413d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3414674ae819SStefano Zampini } 3415d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3416d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3417d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3418d06fc5fdSStefano Zampini } 3419d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3420d06fc5fdSStefano Zampini n_ISForEdges = 0; 3421d06fc5fdSStefano Zampini } 3422d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3423d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3424d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3425d06fc5fdSStefano Zampini } 3426d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3427d06fc5fdSStefano Zampini n_ISForFaces = 0; 3428d06fc5fdSStefano Zampini } 342970022509SStefano Zampini 343070022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 343170022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 343270022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 343370022509SStefano Zampini if (pcbddc->NullSpace) { 343470022509SStefano Zampini PetscBool tbool[2],gbool[2]; 343570022509SStefano Zampini 343670022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3437b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3438d06fc5fdSStefano Zampini if (!ISForEdges) { 3439d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3440d06fc5fdSStefano Zampini } 3441b8ffe317SStefano Zampini } 3442d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3443d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3444b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3445d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3446d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 344798a51de6SStefano Zampini } 344870022509SStefano Zampini #endif 344908122e43SStefano Zampini 3450674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3451674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3452674ae819SStefano Zampini if (nearnullsp) { 3453674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3454f4ddd8eeSStefano Zampini /* remove any stored info */ 3455f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3456f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3457f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3458f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3459f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3460473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3461f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3462f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3463f4ddd8eeSStefano Zampini } 3464984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3465984c4197SStefano Zampini nnsp_size = 0; 3466674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3467674ae819SStefano Zampini } 3468984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3469984c4197SStefano Zampini max_constraints = nnsp_size; 3470984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3471984c4197SStefano Zampini 3472674ae819SStefano Zampini /* 3473674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 34749162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 34759162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 34769162d606SStefano Zampini There can be multiple constraints per connected component 3477674ae819SStefano Zampini */ 3478674ae819SStefano Zampini n_vertices = 0; 3479674ae819SStefano Zampini if (ISForVertices) { 3480674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3481674ae819SStefano Zampini } 34829162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34839162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34849162d606SStefano Zampini 34859162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34869162d606SStefano Zampini total_counts *= max_constraints; 3487674ae819SStefano Zampini total_counts += n_vertices; 34884641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34899162d606SStefano Zampini 3490674ae819SStefano Zampini total_counts = 0; 3491674ae819SStefano Zampini max_size_of_constraint = 0; 3492674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34939162d606SStefano Zampini IS used_is; 3494674ae819SStefano Zampini if (i<n_ISForEdges) { 34959162d606SStefano Zampini used_is = ISForEdges[i]; 3496674ae819SStefano Zampini } else { 34979162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3498674ae819SStefano Zampini } 34999162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3500674ae819SStefano Zampini total_counts += j; 3501674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3502674ae819SStefano Zampini } 35039162d606SStefano 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); 35049162d606SStefano Zampini 3505984c4197SStefano Zampini /* get local part of global near null space vectors */ 3506785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3507984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3508984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3509e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3510e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3511984c4197SStefano Zampini } 3512674ae819SStefano Zampini 3513242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3514242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3515a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3516242a89d7SStefano Zampini 3517984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3518a773dcb8SStefano Zampini if (!skip_lapack) { 3519674ae819SStefano Zampini PetscScalar temp_work; 3520911cabfeSStefano Zampini 3521674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3522984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3523785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3524785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3525785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3526674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3527785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3528674ae819SStefano Zampini #endif 3529674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3530c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3531c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3532674ae819SStefano Zampini lwork = -1; 3533674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3534674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3535c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3536674ae819SStefano Zampini #else 3537c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3538674ae819SStefano Zampini #endif 3539674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3540984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3541674ae819SStefano Zampini #else /* on missing GESVD */ 3542674ae819SStefano Zampini /* SVD */ 3543674ae819SStefano Zampini PetscInt max_n,min_n; 3544674ae819SStefano Zampini max_n = max_size_of_constraint; 3545984c4197SStefano Zampini min_n = max_constraints; 3546984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3547674ae819SStefano Zampini min_n = max_size_of_constraint; 3548984c4197SStefano Zampini max_n = max_constraints; 3549674ae819SStefano Zampini } 3550785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3551674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3552785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3553674ae819SStefano Zampini #endif 3554674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3555674ae819SStefano Zampini lwork = -1; 3556e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3557e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3558b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3559674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3560674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35619162d606SStefano 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)); 3562674ae819SStefano Zampini #else 35639162d606SStefano 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)); 3564674ae819SStefano Zampini #endif 3565674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3566984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3567984c4197SStefano Zampini #endif /* on missing GESVD */ 3568674ae819SStefano Zampini /* Allocate optimal workspace */ 3569674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3570854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3571674ae819SStefano Zampini } 3572674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3573674ae819SStefano Zampini total_counts = 0; 35749162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 35759162d606SStefano Zampini constraints_data_ptr[0] = 0; 3576674ae819SStefano Zampini /* vertices */ 35779162d606SStefano Zampini if (n_vertices) { 3578674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 35799162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3580674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35819162d606SStefano Zampini constraints_n[total_counts] = 1; 35829162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35839162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35849162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3585674ae819SStefano Zampini total_counts++; 3586674ae819SStefano Zampini } 3587674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3588674ae819SStefano Zampini n_vertices = total_counts; 3589674ae819SStefano Zampini } 3590984c4197SStefano Zampini 3591674ae819SStefano Zampini /* edges and faces */ 35929162d606SStefano Zampini total_counts_cc = total_counts; 3593911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35949162d606SStefano Zampini IS used_is; 35959162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35969162d606SStefano Zampini 3597911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35989162d606SStefano Zampini used_is = ISForEdges[ncc]; 3599984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3600674ae819SStefano Zampini } else { 36019162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3602984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3603674ae819SStefano Zampini } 3604674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 36059162d606SStefano Zampini 36069162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 36079162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3608984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3609984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3610674ae819SStefano Zampini if (nnsp_has_cnst) { 36115b08dc53SStefano Zampini PetscScalar quad_value; 36129162d606SStefano Zampini 36139162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 36149162d606SStefano Zampini idxs_copied = PETSC_TRUE; 36159162d606SStefano Zampini 3616a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3617674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3618a773dcb8SStefano Zampini } else { 3619a773dcb8SStefano Zampini quad_value = 1.0; 3620a773dcb8SStefano Zampini } 3621674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 36229162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3623674ae819SStefano Zampini } 36249162d606SStefano Zampini temp_constraints++; 3625674ae819SStefano Zampini total_counts++; 3626674ae819SStefano Zampini } 3627674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3628984c4197SStefano Zampini PetscReal real_value; 36299162d606SStefano Zampini PetscScalar *ptr_to_data; 36309162d606SStefano Zampini 3631984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 36329162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3633674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 36349162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3635674ae819SStefano Zampini } 3636984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3637984c4197SStefano Zampini /* check if array is null on the connected component */ 3638e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 36399162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 36405b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3641674ae819SStefano Zampini temp_constraints++; 3642674ae819SStefano Zampini total_counts++; 36439162d606SStefano Zampini if (!idxs_copied) { 36449162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 36459162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3646674ae819SStefano Zampini } 3647674ae819SStefano Zampini } 36489162d606SStefano Zampini } 36499162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 365045a1bb75SStefano Zampini valid_constraints = temp_constraints; 3651eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3652a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 36539162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 36549162d606SStefano Zampini 36559162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3656a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 36579162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3658a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 36599162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3660a773dcb8SStefano Zampini } else { /* perform SVD */ 3661984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 36629162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3663674ae819SStefano Zampini 3664674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3665984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3666984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3667984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3668984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3669984c4197SStefano Zampini from that computed using LAPACKgesvd 3670984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3671984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3672984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3673674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3674e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3675984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3676674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3677674ae819SStefano Zampini for (k=0;k<j+1;k++) { 36789162d606SStefano 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)); 3679674ae819SStefano Zampini } 3680674ae819SStefano Zampini } 3681e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3682e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3683e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3684674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3685c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3686674ae819SStefano Zampini #else 3687c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3688674ae819SStefano Zampini #endif 3689674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3690984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3691984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3692674ae819SStefano Zampini j = 0; 3693984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3694674ae819SStefano Zampini total_counts = total_counts-j; 369545a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3696e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3697c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3698c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3699c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3700c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3701c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3702c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3703674ae819SStefano Zampini if (j<temp_constraints) { 3704984c4197SStefano Zampini PetscInt ii; 3705984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3706674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 37079162d606SStefano 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)); 3708674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3709984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3710674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 37119162d606SStefano 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]; 3712674ae819SStefano Zampini } 3713674ae819SStefano Zampini } 3714674ae819SStefano Zampini } 3715674ae819SStefano Zampini #else /* on missing GESVD */ 3716e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3717e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3718b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3719674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3720674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 37219162d606SStefano 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)); 3722674ae819SStefano Zampini #else 37239162d606SStefano 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)); 3724674ae819SStefano Zampini #endif 3725984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3726674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3727984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3728e310c8b4SStefano Zampini k = temp_constraints; 3729e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3730674ae819SStefano Zampini j = 0; 3731e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 373245a1bb75SStefano Zampini valid_constraints = k-j; 3733911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3734984c4197SStefano Zampini #endif /* on missing GESVD */ 3735674ae819SStefano Zampini } 3736a773dcb8SStefano Zampini } 37379162d606SStefano Zampini /* update pointers information */ 37389162d606SStefano Zampini if (valid_constraints) { 37399162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 37409162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 37419162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 37429162d606SStefano Zampini /* set change_of_basis flag */ 374345a1bb75SStefano Zampini if (boolforchange) { 3744b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 37459162d606SStefano Zampini } 3746b3d85658SStefano Zampini total_counts_cc++; 374745a1bb75SStefano Zampini } 374845a1bb75SStefano Zampini } 3749984c4197SStefano Zampini /* free workspace */ 37508f1c130eSStefano Zampini if (!skip_lapack) { 3751984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3752984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3753984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3754984c4197SStefano Zampini #endif 3755984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3756984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3757984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3758984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3759984c4197SStefano Zampini #endif 3760984c4197SStefano Zampini } 3761984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3762984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3763984c4197SStefano Zampini } 3764984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3765cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3766cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3767cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3768cf5a6209SStefano Zampini } 3769cf5a6209SStefano Zampini if (n_ISForFaces) { 3770cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3771cf5a6209SStefano Zampini } 3772cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3773cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3774cf5a6209SStefano Zampini } 3775cf5a6209SStefano Zampini if (n_ISForEdges) { 3776cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3777cf5a6209SStefano Zampini } 3778cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 377908122e43SStefano Zampini } else { 378008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3781984c4197SStefano Zampini 378208122e43SStefano Zampini total_counts = 0; 378308122e43SStefano Zampini n_vertices = 0; 3784d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3785d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 378608122e43SStefano Zampini } 378708122e43SStefano Zampini max_constraints = 0; 37889162d606SStefano Zampini total_counts_cc = 0; 378908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 379008122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37919162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 379208122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 379308122e43SStefano Zampini } 37949162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37959162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37969162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37979162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 379874d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37999162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 38009162d606SStefano Zampini total_counts_cc = 0; 38019162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 38029162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 38039162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 380408122e43SStefano Zampini } 380508122e43SStefano Zampini } 38069162d606SStefano Zampini #if 0 38079162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 38089162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 38099162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 38109162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 38119162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 38129162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 38139162d606SStefano Zampini } 38149162d606SStefano Zampini printf("\n"); 38159162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 38169162d606SStefano Zampini } 38171b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 38188bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 38191b968477SStefano Zampini } 38201b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 38218bec7fa6SStefano 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]); 38221b968477SStefano Zampini } 382308122e43SStefano Zampini #endif 382408122e43SStefano Zampini 38258bec7fa6SStefano Zampini max_size_of_constraint = 0; 38269162d606SStefano 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]); 38279162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 382808122e43SStefano Zampini /* Change of basis */ 3829b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 383008122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 383108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 383208122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3833b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 383408122e43SStefano Zampini } 383508122e43SStefano Zampini } 383608122e43SStefano Zampini } 383708122e43SStefano Zampini } 3838984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 38394f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 384008122e43SStefano Zampini 38419162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 38429162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 38436c4ed002SBarry 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); 3844674ae819SStefano Zampini 3845674ae819SStefano Zampini /* Create constraint matrix */ 3846674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 384716f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3848984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3849984c4197SStefano Zampini 3850984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3851a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3852a717540cSStefano Zampini qr_needed = PETSC_FALSE; 385374d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3854984c4197SStefano Zampini total_primal_vertices=0; 3855b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 38569162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 38579162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 385872b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 38599162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3860b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 386164efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 38629162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38639162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3864a717540cSStefano Zampini } 3865b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 386691af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 3867a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3868a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3869a717540cSStefano Zampini } 3870fa434743SStefano Zampini } else { 3871b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3872fa434743SStefano Zampini } 3873a717540cSStefano Zampini } 3874b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3875b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3876674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 387770022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3878b3d85658SStefano Zampini 38794f1b2e48SStefano 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); 38800e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38810e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3882984c4197SStefano Zampini 3883984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 388474d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3885785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3886984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 388774d5cdf7SStefano Zampini 3888984c4197SStefano Zampini j = total_primal_vertices; 388974d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3890b3d85658SStefano Zampini cum = total_primal_vertices; 38919162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38924641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3893b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3894b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3895b3d85658SStefano Zampini cum++; 38969162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 389774d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 389874d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 389974d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 390074d5cdf7SStefano Zampini } 39019162d606SStefano Zampini j += constraints_n[i]; 3902674ae819SStefano Zampini } 3903674ae819SStefano Zampini } 3904674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3905674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3906088faed8SStefano Zampini 3907674ae819SStefano Zampini /* set values in constraint matrix */ 3908984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 39090e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3910674ae819SStefano Zampini } 3911984c4197SStefano Zampini total_counts = total_primal_vertices; 39129162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 39134641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 39149162d606SStefano Zampini PetscInt *cols; 39159162d606SStefano Zampini 39169162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 39179162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 39189162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 39199162d606SStefano Zampini PetscInt row = total_counts+k; 39209162d606SStefano Zampini PetscScalar *vals; 39219162d606SStefano Zampini 39229162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 39239162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 39249162d606SStefano Zampini } 39259162d606SStefano Zampini total_counts += constraints_n[i]; 3926674ae819SStefano Zampini } 3927674ae819SStefano Zampini } 3928674ae819SStefano Zampini /* assembling */ 3929674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3930674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3931088faed8SStefano Zampini 3932984c4197SStefano Zampini /* 39336a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3934984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3935f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 3936984c4197SStefano Zampini */ 3937674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3938674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3939026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3940984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3941984c4197SStefano Zampini /* working stuff for GEQRF */ 394281d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3943984c4197SStefano Zampini PetscBLASInt lqr_work; 3944984c4197SStefano Zampini /* working stuff for UNGQR */ 3945984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3946984c4197SStefano Zampini PetscBLASInt lgqr_work; 3947984c4197SStefano Zampini /* working stuff for TRTRS */ 3948984c4197SStefano Zampini PetscScalar *trs_rhs; 39493f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3950984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3951984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3952984c4197SStefano Zampini PetscScalar *start_vals; 3953984c4197SStefano Zampini /* working stuff for values insertion */ 39544641a718SStefano Zampini PetscBT is_primal; 395564efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3956906d46d4SStefano Zampini /* matrix sizes */ 3957906d46d4SStefano Zampini PetscInt global_size,local_size; 3958906d46d4SStefano Zampini /* temporary change of basis */ 3959906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3960cf5a6209SStefano Zampini /* extra space for debugging */ 3961cf5a6209SStefano Zampini PetscScalar *dbg_work; 3962984c4197SStefano Zampini 3963906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3964906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 396516f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3966bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3967906d46d4SStefano Zampini /* nonzeros for local mat */ 3968bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 39691dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3970bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 39711dd7afcfSStefano Zampini } else { 39721dd7afcfSStefano Zampini const PetscInt *ii; 39731dd7afcfSStefano Zampini PetscInt n; 39741dd7afcfSStefano Zampini PetscBool flg_row; 39751dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39761dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 39771dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39781dd7afcfSStefano Zampini } 39799162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3980a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39819162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3982a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39839162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3984a717540cSStefano Zampini } else { 39859162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39869162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3987a717540cSStefano Zampini } 3988a717540cSStefano Zampini } 3989a717540cSStefano Zampini } 3990906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3991bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 39921dd7afcfSStefano Zampini /* Set interior change in the matrix */ 39931dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3994bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3995906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3996a717540cSStefano Zampini } 39971dd7afcfSStefano Zampini } else { 39981dd7afcfSStefano Zampini const PetscInt *ii,*jj; 39991dd7afcfSStefano Zampini PetscScalar *aa; 40001dd7afcfSStefano Zampini PetscInt n; 40011dd7afcfSStefano Zampini PetscBool flg_row; 40021dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 40031dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 40041dd7afcfSStefano Zampini for (i=0;i<n;i++) { 40051dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 40061dd7afcfSStefano Zampini } 40071dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 40081dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 40091dd7afcfSStefano Zampini } 4010a717540cSStefano Zampini 4011a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4012a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 4013a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 4014a717540cSStefano Zampini } 4015a717540cSStefano Zampini 4016a717540cSStefano Zampini 4017a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 4018a717540cSStefano Zampini /* 4019a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 4020a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 4021a717540cSStefano Zampini 4022a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 4023a717540cSStefano Zampini 4024a6b551f4SStefano 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) 4025a6b551f4SStefano Zampini 4026a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 4027a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 4028a717540cSStefano Zampini | ... | 4029a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 4030a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 4031a717540cSStefano Zampini 4032a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 4033a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 4034a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 4035a6b551f4SStefano Zampini 4036a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 4037a717540cSStefano Zampini */ 4038a717540cSStefano Zampini if (qr_needed) { 4039984c4197SStefano Zampini /* space to store Q */ 4040854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 4041984c4197SStefano Zampini /* first we issue queries for optimal work */ 40423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 40433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 40443f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4045984c4197SStefano Zampini lqr_work = -1; 40463f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4047984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4048984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4049785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4050984c4197SStefano Zampini lgqr_work = -1; 40513f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 40523f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 40533f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 40543f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40553f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 40563f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4057984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4058984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4059785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4060984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4061785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4062984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4063785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4064a717540cSStefano Zampini /* allocating workspace for check */ 4065a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4066cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4067a717540cSStefano Zampini } 4068a717540cSStefano Zampini } 4069984c4197SStefano Zampini /* array to store whether a node is primal or not */ 40704641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4071473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 40720e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 40736c4ed002SBarry 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); 407439e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 407539e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 407639e2fb2aSStefano Zampini } 407739e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4078984c4197SStefano Zampini 4079a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 40809162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40819162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40824641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4083984c4197SStefano Zampini /* get constraint info */ 40849162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4085984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4086984c4197SStefano Zampini 4087984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40889162d606SStefano 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); 4089674ae819SStefano Zampini } 4090984c4197SStefano Zampini 4091fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4092a717540cSStefano Zampini 4093a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4094a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40959162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4096a717540cSStefano Zampini } 4097984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40989162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4099984c4197SStefano Zampini 4100984c4197SStefano Zampini /* compute QR decomposition of constraints */ 41013f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 41023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 41033f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4104674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 41053f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4106984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4107674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4108984c4197SStefano Zampini 4109984c4197SStefano Zampini /* explictly compute R^-T */ 4110984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4111984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 41123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 41133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 41143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 41153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4116984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 41173f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4118984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4119984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4120984c4197SStefano Zampini 4121a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 41223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 41233f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41243f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 41253f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4126984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 41273f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4128984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4129984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4130984c4197SStefano Zampini 4131984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4132984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4133984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 41343f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 41353f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 41363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 41373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 41383f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 41393f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4140984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 41419162d606SStefano 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)); 4142984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 41439162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4144984c4197SStefano Zampini 4145984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 41469162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4147984c4197SStefano Zampini /* insert cols for primal dofs */ 4148984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4149984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 41509162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4151906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4152984c4197SStefano Zampini } 4153984c4197SStefano Zampini /* insert cols for dual dofs */ 4154984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 41559162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4156984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 41579162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4158906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4159984c4197SStefano Zampini j++; 4160674ae819SStefano Zampini } 4161674ae819SStefano Zampini } 4162984c4197SStefano Zampini 4163984c4197SStefano Zampini /* check change of basis */ 4164984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4165984c4197SStefano Zampini PetscInt ii,jj; 4166984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4167c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4168c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4169c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4170c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4171c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4172c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4173984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4174cf5a6209SStefano 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)); 4175984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4176984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4177984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4178cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4179cf5a6209SStefano 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; 4180674ae819SStefano Zampini } 4181674ae819SStefano Zampini } 4182984c4197SStefano Zampini if (!valid_qr) { 418322d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4184984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4185984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4186cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4187cf5a6209SStefano 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])); 4188674ae819SStefano Zampini } 4189cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4190cf5a6209SStefano 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])); 4191984c4197SStefano Zampini } 4192984c4197SStefano Zampini } 4193984c4197SStefano Zampini } 4194674ae819SStefano Zampini } else { 419522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4196674ae819SStefano Zampini } 4197674ae819SStefano Zampini } 4198a717540cSStefano Zampini } else { /* simple transformation block */ 4199a717540cSStefano Zampini PetscInt row,col; 4200a6b551f4SStefano Zampini PetscScalar val,norm; 4201a6b551f4SStefano Zampini 4202a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 42039162d606SStefano 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)); 4204a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 42059162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 42069162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4207bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 42089162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4209906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 42109162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4211a717540cSStefano Zampini } else { 4212a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 42139162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4214a717540cSStefano Zampini if (row != col) { 42159162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4216a717540cSStefano Zampini } else { 42179162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4218a717540cSStefano Zampini } 4219906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4220a717540cSStefano Zampini } 4221a717540cSStefano Zampini } 4222a717540cSStefano Zampini } 422398a51de6SStefano Zampini if (pcbddc->dbg_flag) { 422422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4225a717540cSStefano Zampini } 4226674ae819SStefano Zampini } 4227984c4197SStefano Zampini } else { 4228984c4197SStefano Zampini if (pcbddc->dbg_flag) { 42299162d606SStefano 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); 4230674ae819SStefano Zampini } 4231674ae819SStefano Zampini } 4232674ae819SStefano Zampini } 4233a717540cSStefano Zampini 4234a717540cSStefano Zampini /* free workspace */ 4235a717540cSStefano Zampini if (qr_needed) { 4236984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4237cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4238984c4197SStefano Zampini } 4239984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4240984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4241984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4242984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4243984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4244674ae819SStefano Zampini } 4245a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4246906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4247906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4248906d46d4SStefano Zampini 4249906d46d4SStefano Zampini /* assembling of global change of variable */ 425088c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4251bbb9e6c6SStefano Zampini Mat tmat; 425216f15bc4SStefano Zampini PetscInt bs; 425316f15bc4SStefano Zampini 4254906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4255906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4256bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4257bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4258bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4259bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 426016f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 426116f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4262906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4263bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4264bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4265bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4266bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4267bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4268e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4269e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4270bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4271bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 427288c03ad3SStefano Zampini 4273906d46d4SStefano Zampini /* check */ 4274906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4275906d46d4SStefano Zampini PetscReal error; 4276906d46d4SStefano Zampini Vec x,x_change; 4277906d46d4SStefano Zampini 4278906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4279906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4280906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4281906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4282e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4283e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4284bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4285e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4286e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4287906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4288906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4289906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4290906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4291bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4292906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4293906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4294906d46d4SStefano Zampini } 4295b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4296b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4297b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4298bf3a8328SStefano Zampini 4299bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4300bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4301bf3a8328SStefano Zampini } 4302b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4303ac632422SStefano Zampini Mat S_new,tmat; 4304bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4305bbb9e6c6SStefano Zampini 4306bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 43076816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4308bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4309bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4310bf3a8328SStefano Zampini IS is_V; 4311b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4312b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4313b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4314b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4315b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4316bf3a8328SStefano Zampini } 4317bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4318ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4319b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4320ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4321bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4322bf3a8328SStefano Zampini const PetscScalar *array; 4323bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4324bf3a8328SStefano Zampini PetscInt i,n_V; 4325bf3a8328SStefano Zampini 4326b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4327b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4328b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4329b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4330b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4331b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4332b087196eSStefano Zampini PetscScalar val; 4333b087196eSStefano Zampini PetscInt idx; 4334b087196eSStefano Zampini 4335b087196eSStefano Zampini idx = idxs_V[i]; 4336b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4337b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4338b087196eSStefano Zampini } 4339b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4340b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4341bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4342bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4343bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4344bf3a8328SStefano Zampini } 4345ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4346ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4347ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4348ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4349b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4350ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4351bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4352b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4353bf3a8328SStefano Zampini } 4354ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4355ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4356ac632422SStefano Zampini } 4357b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 435888c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4359b96c3477SStefano Zampini } 4360c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4361c9db6a07SStefano Zampini if (sub_schurs->change) { 4362c9db6a07SStefano Zampini PetscInt i; 4363c9db6a07SStefano Zampini 4364c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4365c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4366c9db6a07SStefano Zampini } 4367c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4368c9db6a07SStefano Zampini } 4369b96c3477SStefano Zampini } 437016909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 437116909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 437216909a7fSStefano Zampini } else { 4373906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 437416909a7fSStefano Zampini } 43751dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 437627b6a85dSStefano Zampini pcbddc->change_interior = pcbddc->benign_have_null; 437772b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 437872b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 437972b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 438072b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 438172b8c272SStefano Zampini } 43821dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 438327b6a85dSStefano Zampini if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) { 4384b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4385b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4386906d46d4SStefano Zampini } else { 43871dd7afcfSStefano Zampini Mat benign_global = NULL; 438827b6a85dSStefano Zampini if (pcbddc->benign_have_null) { 43891dd7afcfSStefano Zampini Mat tmat; 43901dd7afcfSStefano Zampini 43911dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 43921dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 43931dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 43941dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43951dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43961dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 43971dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43981dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43991dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 44001dd7afcfSStefano Zampini if (pcbddc->benign_change) { 44011dd7afcfSStefano Zampini Mat M; 44021dd7afcfSStefano Zampini 44031dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 44041dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 44051dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 44061dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4407906d46d4SStefano Zampini } else { 44081dd7afcfSStefano Zampini Mat eye; 44091dd7afcfSStefano Zampini PetscScalar *array; 44101dd7afcfSStefano Zampini 44111dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 44121dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 44131dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 44141dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4415906d46d4SStefano Zampini } 44161dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 44171dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 44181dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 44191dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 44201dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 44211dd7afcfSStefano Zampini } 44221dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 44231dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 44241dd7afcfSStefano Zampini } 44251dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 44261dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 44271dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 442827b6a85dSStefano Zampini } else if (pcbddc->benign_have_null) { 44291dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 44301dd7afcfSStefano Zampini } 44311dd7afcfSStefano Zampini } 443216909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 443316909a7fSStefano Zampini IS is_global; 443416909a7fSStefano Zampini const PetscInt *gidxs; 443516909a7fSStefano Zampini 443616909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 443716909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 443816909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 443916909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 444016909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 444116909a7fSStefano Zampini } 44421dd7afcfSStefano Zampini } 44431dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 44441dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4445b9b85e73SStefano Zampini } 4446a717540cSStefano Zampini 444772b8c272SStefano Zampini if (!pcbddc->fake_change) { 44484f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 44494f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 44504f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 44514f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4452019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4453019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4454019a44ceSStefano Zampini pcbddc->local_primal_size++; 4455019a44ceSStefano Zampini } 4456019a44ceSStefano Zampini 4457019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4458727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4459727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 44609f47a83aSStefano 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); 4461c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 44620e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 44639f47a83aSStefano 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); 4464727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4465727cdba6SStefano Zampini } 44660e6343abSStefano Zampini } 4467727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4468b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 446972b8c272SStefano Zampini } 447072b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4471727cdba6SStefano Zampini 4472a717540cSStefano Zampini /* flush dbg viewer */ 4473b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4474b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4475b8ffe317SStefano Zampini } 4476a717540cSStefano Zampini 4477e310c8b4SStefano Zampini /* free workspace */ 4478a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 44794641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 448008122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 44819162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 44829162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 448308122e43SStefano Zampini } else { 44849162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 44859162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 44869162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 448708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 448808122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 44899162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 44909162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 449108122e43SStefano Zampini } 4492674ae819SStefano Zampini PetscFunctionReturn(0); 4493674ae819SStefano Zampini } 4494674ae819SStefano Zampini 4495674ae819SStefano Zampini #undef __FUNCT__ 4496674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4497674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4498674ae819SStefano Zampini { 4499674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4500674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4501674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 450214f95afaSStefano Zampini PetscInt ierr,i,N; 4503674ae819SStefano Zampini 4504674ae819SStefano Zampini PetscFunctionBegin; 45058e61c736SStefano Zampini /* Reset previously computed graph */ 45068e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4507674ae819SStefano Zampini /* Init local Graph struct */ 45087fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 45093bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4510674ae819SStefano Zampini 4511575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 45125099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 45135099eff2SStefano 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); 4514575ad6abSStefano Zampini } 45159577ea80SStefano Zampini 4516674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4517d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 45184d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 45194d379d7bSStefano Zampini PetscInt nvtxs; 4520e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4521674ae819SStefano Zampini 45222fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 45232fffb893SStefano Zampini if (flg_row) { 45244d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4525b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 45262fffb893SStefano Zampini } 45272fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4528674ae819SStefano Zampini } 45299b28b941SStefano Zampini if (pcbddc->dbg_flag) { 45309b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4531674ae819SStefano Zampini } 4532674ae819SStefano Zampini 453363602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 453463602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 453563602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 453695ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 453763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4538e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 453963602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4540674ae819SStefano Zampini } 454163602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 454263602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 454363602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4544674ae819SStefano Zampini } 454563602bcaSStefano Zampini } else { 454663602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 454763602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4548854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 454963602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 455063602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 455163602bcaSStefano Zampini } 455263602bcaSStefano Zampini } 4553674ae819SStefano Zampini } 4554674ae819SStefano Zampini 4555674ae819SStefano Zampini /* Setup of Graph */ 4556785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4557e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4558785d1243SStefano Zampini } 4559785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4560e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4561785d1243SStefano Zampini } 456230368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 456330368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 456430368db7SStefano Zampini } 456514f95afaSStefano 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); 4566674ae819SStefano Zampini 45674f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 45684f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 45694f1b2e48SStefano Zampini PetscInt *local_subs; 45704f1b2e48SStefano Zampini 45714f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 45724f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 45734f1b2e48SStefano Zampini const PetscInt *idxs; 45744f1b2e48SStefano Zampini PetscInt nl,j; 45754f1b2e48SStefano Zampini 45764f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 45774f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45784f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 45794f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 45804f1b2e48SStefano Zampini } 45814f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45824f1b2e48SStefano Zampini } 45834f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 45844f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 45854f1b2e48SStefano Zampini } 45864f1b2e48SStefano Zampini 4587674ae819SStefano Zampini /* Graph's connected components analysis */ 4588674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4589674ae819SStefano Zampini PetscFunctionReturn(0); 4590674ae819SStefano Zampini } 4591674ae819SStefano Zampini 4592dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4593674ae819SStefano Zampini #undef __FUNCT__ 4594674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4595dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4596674ae819SStefano Zampini { 4597dc456d91SStefano Zampini PetscSF sf; 4598dc456d91SStefano Zampini PetscLayout map; 4599dc456d91SStefano Zampini const PetscInt *idxs; 4600dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4601dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4602dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4603dc456d91SStefano Zampini PetscMPIInt commsize; 4604674ae819SStefano Zampini PetscBool first_found; 4605674ae819SStefano Zampini PetscErrorCode ierr; 4606674ae819SStefano Zampini 4607674ae819SStefano Zampini PetscFunctionBegin; 4608dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4609dc456d91SStefano Zampini if (subset_mult) { 4610dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4611dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4612dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4613674ae819SStefano Zampini } 4614dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4615dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4616dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4617dc456d91SStefano Zampini for (i=0;i<n;i++) { 4618dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4619dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4620674ae819SStefano Zampini } 4621dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4622b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4623dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4624dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4625dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4626dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4627dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4628dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4629dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4630dc456d91SStefano Zampini 4631dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4632dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4633dc456d91SStefano Zampini if (subset_mult) { 4634dc456d91SStefano Zampini const PetscInt* idxs_mult; 4635dc456d91SStefano Zampini 4636dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4637dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4638dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4639674ae819SStefano Zampini } else { 4640dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4641674ae819SStefano Zampini } 4642dc456d91SStefano Zampini /* local size of new subset */ 4643dc456d91SStefano Zampini n_n = 0; 4644dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4645dc456d91SStefano Zampini 4646dc456d91SStefano Zampini /* global indexes in layout */ 4647dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4648dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4649dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4650dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4651dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4652dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4653dc456d91SStefano Zampini 4654dc456d91SStefano Zampini /* reduce from leaves to roots */ 4655dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 465664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 465764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4658dc456d91SStefano Zampini 4659dc456d91SStefano Zampini /* count indexes in local part of layout */ 4660674ae819SStefano Zampini nlocals = 0; 4661674ae819SStefano Zampini first_index = -1; 4662674ae819SStefano Zampini first_found = PETSC_FALSE; 4663dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4664dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4665674ae819SStefano Zampini first_found = PETSC_TRUE; 4666674ae819SStefano Zampini first_index = i; 4667674ae819SStefano Zampini } 4668dc456d91SStefano Zampini nlocals += root_data[i]; 4669674ae819SStefano Zampini } 4670dc456d91SStefano Zampini 4671dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 46725fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4673dc456d91SStefano Zampini start = 0; 467464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46755fa240b1SStefano Zampini #else 467664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46775fa240b1SStefano Zampini start = start-nlocals; 46785fa240b1SStefano Zampini #endif 46795fa240b1SStefano Zampini 4680dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4681dc456d91SStefano Zampini *N_n = start + nlocals; 4682dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4683dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4684674ae819SStefano Zampini } 46855fa240b1SStefano Zampini 46865fa240b1SStefano Zampini /* adapt root data with cumulative */ 4687674ae819SStefano Zampini if (first_found) { 4688dc456d91SStefano Zampini PetscInt old_index; 4689dc456d91SStefano Zampini 4690dc456d91SStefano Zampini root_data[first_index] += start; 4691674ae819SStefano Zampini old_index = first_index; 4692dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4693dc456d91SStefano Zampini if (root_data[i]) { 4694dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4695674ae819SStefano Zampini old_index = i; 4696674ae819SStefano Zampini } 4697674ae819SStefano Zampini } 4698674ae819SStefano Zampini } 4699dc456d91SStefano Zampini 4700dc456d91SStefano Zampini /* from roots to leaves */ 4701dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4702dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4703dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4704dc456d91SStefano Zampini 4705dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4706dc456d91SStefano Zampini if (subset_mult) { 4707dc456d91SStefano Zampini const PetscInt* idxs_mult; 4708dc456d91SStefano Zampini PetscInt cum; 4709dc456d91SStefano Zampini 4710dc456d91SStefano Zampini cum = 0; 4711dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4712dc456d91SStefano Zampini for (i=0;i<n;i++) { 4713dc456d91SStefano Zampini PetscInt j; 4714dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4715674ae819SStefano Zampini } 4716dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4717674ae819SStefano Zampini } else { 4718dc456d91SStefano Zampini for (i=0;i<n;i++) { 4719dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4720674ae819SStefano Zampini } 4721674ae819SStefano Zampini } 4722dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4723dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4724674ae819SStefano Zampini PetscFunctionReturn(0); 4725674ae819SStefano Zampini } 47269a7d3425SStefano Zampini 47279a7d3425SStefano Zampini #undef __FUNCT__ 47289a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 47299a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 47309a7d3425SStefano Zampini { 47319a7d3425SStefano Zampini PetscInt i,j; 47329a7d3425SStefano Zampini PetscScalar *alphas; 47339a7d3425SStefano Zampini PetscErrorCode ierr; 47349a7d3425SStefano Zampini 47359a7d3425SStefano Zampini PetscFunctionBegin; 47369a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4737785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 47389a7d3425SStefano Zampini for (i=0;i<n;i++) { 47399a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 47409a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 47419a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 47429a7d3425SStefano Zampini } 47439a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 47449a7d3425SStefano Zampini PetscFunctionReturn(0); 47459a7d3425SStefano Zampini } 47469a7d3425SStefano Zampini 4747e7931f94SStefano Zampini #undef __FUNCT__ 474870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 474957de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4750e7931f94SStefano Zampini { 475157de7509SStefano Zampini Mat A; 4752e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4753e7931f94SStefano Zampini PetscMPIInt size,rank,color; 475452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 475552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 475627b6a85dSStefano Zampini PetscInt im_active,active_procs,n,i,j,local_size,threshold = 2; 475757de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 475827b6a85dSStefano Zampini PetscInt xadj_count, *count; 475927b6a85dSStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE; 476027b6a85dSStefano Zampini PetscSubcomm psubcomm; 476127b6a85dSStefano Zampini MPI_Comm subcomm; 476252e5ac9dSStefano Zampini PetscErrorCode ierr; 4763a57a6d2fSStefano Zampini 4764e7931f94SStefano Zampini PetscFunctionBegin; 476557de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 476657de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 476757de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 476857de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 476957de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 477057de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 477157de7509SStefano Zampini 477257de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 477357de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 477457de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 477557de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 477657de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 477757de7509SStefano Zampini im_active = !!(n); 477857de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 477957de7509SStefano Zampini void_procs = size - active_procs; 478057de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 478157de7509SStefano Zampini if (void_procs) { 478257de7509SStefano Zampini PetscInt ncand; 478357de7509SStefano Zampini 478457de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 478557de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 478657de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 478757de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 478857de7509SStefano Zampini if (!procs_candidates[i]) { 478957de7509SStefano Zampini procs_candidates[ncand++] = i; 479057de7509SStefano Zampini } 479157de7509SStefano Zampini } 479257de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 479357de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 479457de7509SStefano Zampini } 479557de7509SStefano Zampini 479657de7509SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix */ 479757de7509SStefano Zampini if (active_procs < *n_subdomains) { 479857de7509SStefano Zampini PetscInt issize,isidx; 479957de7509SStefano Zampini if (im_active) { 480057de7509SStefano Zampini issize = 1; 480157de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 480257de7509SStefano Zampini isidx = procs_candidates[rank]; 480357de7509SStefano Zampini } else { 480457de7509SStefano Zampini isidx = rank; 480557de7509SStefano Zampini } 480657de7509SStefano Zampini } else { 480757de7509SStefano Zampini issize = 0; 480857de7509SStefano Zampini isidx = -1; 480957de7509SStefano Zampini } 481057de7509SStefano Zampini *n_subdomains = active_procs; 481157de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 4812daf8a457SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 481357de7509SStefano Zampini PetscFunctionReturn(0); 481457de7509SStefano Zampini } 4815c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4816c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 481727b6a85dSStefano Zampini threshold = PetscMax(threshold,2); 4818e7931f94SStefano Zampini 4819e7931f94SStefano Zampini /* Get info on mapping */ 48203bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 48213bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4822e7931f94SStefano Zampini 4823e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4824785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4825e7931f94SStefano Zampini xadj[0] = 0; 4826e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4827785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4828785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 482927b6a85dSStefano Zampini ierr = PetscCalloc1(local_size,&count);CHKERRQ(ierr); 483027b6a85dSStefano Zampini for (i=1;i<n_neighs;i++) 483127b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) 483227b6a85dSStefano Zampini count[shared[i][j]] += 1; 4833e7931f94SStefano Zampini 483427b6a85dSStefano Zampini xadj_count = 0; 48352b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 483627b6a85dSStefano Zampini for (j=0;j<n_shared[i];j++) { 483727b6a85dSStefano Zampini if (count[shared[i][j]] < threshold) { 4838d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4839d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4840d023bfaeSStefano Zampini xadj_count++; 484127b6a85dSStefano Zampini break; 484227b6a85dSStefano Zampini } 4843e7931f94SStefano Zampini } 4844e7931f94SStefano Zampini } 4845d023bfaeSStefano Zampini xadj[1] = xadj_count; 484627b6a85dSStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 48473bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4848e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4849e7931f94SStefano Zampini 48503837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4851e7931f94SStefano Zampini 485227b6a85dSStefano Zampini /* Restrict work on active processes only */ 485327b6a85dSStefano Zampini ierr = PetscMPIIntCast(im_active,&color);CHKERRQ(ierr); 485427b6a85dSStefano Zampini if (void_procs) { 485527b6a85dSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);CHKERRQ(ierr); 485627b6a85dSStefano Zampini ierr = PetscSubcommSetNumber(psubcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 485727b6a85dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(psubcomm,color,rank);CHKERRQ(ierr); 485827b6a85dSStefano Zampini subcomm = PetscSubcommChild(psubcomm); 485927b6a85dSStefano Zampini } else { 486027b6a85dSStefano Zampini psubcomm = NULL; 486127b6a85dSStefano Zampini subcomm = PetscObjectComm((PetscObject)mat); 486227b6a85dSStefano Zampini } 486327b6a85dSStefano Zampini 486427b6a85dSStefano Zampini v_wgt = NULL; 486527b6a85dSStefano Zampini if (!color) { 4866e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4867e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4868e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4869c8587f34SStefano Zampini } else { 487052e5ac9dSStefano Zampini Mat subdomain_adj; 487152e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 487252e5ac9dSStefano Zampini MatPartitioning partitioner; 487327b6a85dSStefano Zampini PetscInt rstart=0,rend=0; 487452e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 487557de7509SStefano Zampini PetscMPIInt size; 4876b0c7d250SStefano Zampini PetscBool aggregate; 4877b0c7d250SStefano Zampini 487827b6a85dSStefano Zampini ierr = MPI_Comm_size(subcomm,&size);CHKERRQ(ierr); 487927b6a85dSStefano Zampini if (void_procs) { 488027b6a85dSStefano Zampini PetscInt prank = rank; 4881785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 488227b6a85dSStefano Zampini ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);CHKERRQ(ierr); 4883e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4884e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4885c8587f34SStefano Zampini } 4886e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 488727b6a85dSStefano Zampini } else { 488827b6a85dSStefano Zampini oldranks = NULL; 488927b6a85dSStefano Zampini } 4890b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 489127b6a85dSStefano Zampini if (aggregate) { /* TODO: all this part could be made more efficient */ 4892b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4893b0c7d250SStefano Zampini PetscMPIInt nrank; 4894b0c7d250SStefano Zampini PetscScalar *vals; 4895b0c7d250SStefano Zampini 489627b6a85dSStefano Zampini ierr = MPI_Comm_rank(subcomm,&nrank);CHKERRQ(ierr); 4897b0c7d250SStefano Zampini lrows = 0; 4898b0c7d250SStefano Zampini if (nrank<redprocs) { 4899b0c7d250SStefano Zampini lrows = size/redprocs; 4900b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4901b0c7d250SStefano Zampini } 490227b6a85dSStefano Zampini ierr = MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4903b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4904b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4905b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4906b0c7d250SStefano Zampini row = nrank; 4907b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4908b0c7d250SStefano Zampini cols = adjncy; 4909b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4910b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4911b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4912b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4913b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 491452e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 491552e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 491652e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4917b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 491827b6a85dSStefano Zampini if (use_vwgt) { 491927b6a85dSStefano Zampini Vec v; 492027b6a85dSStefano Zampini const PetscScalar *array; 492127b6a85dSStefano Zampini PetscInt nl; 492227b6a85dSStefano Zampini 492327b6a85dSStefano Zampini ierr = MatCreateVecs(subdomain_adj,&v,NULL);CHKERRQ(ierr); 492427b6a85dSStefano Zampini ierr = VecSetValue(v,row,(PetscScalar)local_size,INSERT_VALUES);CHKERRQ(ierr); 492527b6a85dSStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 492627b6a85dSStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 492727b6a85dSStefano Zampini ierr = VecGetLocalSize(v,&nl);CHKERRQ(ierr); 492827b6a85dSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 492927b6a85dSStefano Zampini ierr = PetscMalloc1(nl,&v_wgt);CHKERRQ(ierr); 493022db5ddcSStefano Zampini for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]); 493127b6a85dSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 493227b6a85dSStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 493327b6a85dSStefano Zampini } 4934b0c7d250SStefano Zampini } else { 493527b6a85dSStefano Zampini ierr = MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 493627b6a85dSStefano Zampini if (use_vwgt) { 493727b6a85dSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 493827b6a85dSStefano Zampini v_wgt[0] = local_size; 493927b6a85dSStefano Zampini } 4940b0c7d250SStefano Zampini } 494122b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4942e7931f94SStefano Zampini 4943e7931f94SStefano Zampini /* Partition */ 494427b6a85dSStefano Zampini ierr = MatPartitioningCreate(subcomm,&partitioner);CHKERRQ(ierr); 4945e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 494627b6a85dSStefano Zampini if (v_wgt) { 4947e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4948c8587f34SStefano Zampini } 494957de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 495057de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 4951e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4952e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 495322b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4954e7931f94SStefano Zampini 495552e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 495652e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 495752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 495852e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 495957de7509SStefano Zampini if (!aggregate) { 496057de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 496127b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 496227b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 496327b6a85dSStefano Zampini #endif 496457de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 496527b6a85dSStefano Zampini } else if (oldranks) { 4966b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 496727b6a85dSStefano Zampini } else { 496827b6a85dSStefano Zampini ranks_send_to_idx[0] = is_indices[0]; 496957de7509SStefano Zampini } 497028143c3dSStefano Zampini } else { 4971b0c7d250SStefano Zampini PetscInt idxs[1]; 4972b0c7d250SStefano Zampini PetscMPIInt tag; 4973b0c7d250SStefano Zampini MPI_Request *reqs; 4974b0c7d250SStefano Zampini 4975b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4976b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4977b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 497827b6a85dSStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);CHKERRQ(ierr); 497928143c3dSStefano Zampini } 498027b6a85dSStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);CHKERRQ(ierr); 4981b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4982b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 498357de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 498427b6a85dSStefano Zampini #if defined(PETSC_USE_DEBUG) 498527b6a85dSStefano Zampini if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 498627b6a85dSStefano Zampini #endif 498757de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 498827b6a85dSStefano Zampini } else if (oldranks) { 4989b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 499027b6a85dSStefano Zampini } else { 499127b6a85dSStefano Zampini ranks_send_to_idx[0] = idxs[0]; 4992e7931f94SStefano Zampini } 499357de7509SStefano Zampini } 499452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4995e7931f94SStefano Zampini /* clean up */ 4996e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 499752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4998e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4999e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 5000e7931f94SStefano Zampini } 500127b6a85dSStefano Zampini ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 500257de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 5003e7931f94SStefano Zampini 5004e7931f94SStefano Zampini /* assemble parallel IS for sends */ 5005e7931f94SStefano Zampini i = 1; 500627b6a85dSStefano Zampini if (!color) i=0; 500757de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 5008e7931f94SStefano Zampini PetscFunctionReturn(0); 5009e7931f94SStefano Zampini } 5010e7931f94SStefano Zampini 5011e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 5012e7931f94SStefano Zampini 5013e7931f94SStefano Zampini #undef __FUNCT__ 5014e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 501557de7509SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, PetscBool reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 5016e7931f94SStefano Zampini { 501770cf5478SStefano Zampini Mat local_mat; 5018e7931f94SStefano Zampini IS is_sends_internal; 50199d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 502028143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 50219d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 5022e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 5023e7931f94SStefano Zampini PetscInt* l2gmap_indices; 5024e7931f94SStefano Zampini const PetscInt* is_indices; 5025e7931f94SStefano Zampini MatType new_local_type; 5026e7931f94SStefano Zampini /* buffers */ 5027e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 502828143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 50299d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 5030e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 5031e7931f94SStefano Zampini /* MPI */ 503228143c3dSStefano Zampini MPI_Comm comm,comm_n; 503328143c3dSStefano Zampini PetscSubcomm subcomm; 5034e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 503528143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 503628143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 503728143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 503828143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 503928143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 5040e7931f94SStefano Zampini PetscErrorCode ierr; 5041e7931f94SStefano Zampini 5042e7931f94SStefano Zampini PetscFunctionBegin; 504357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5044e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 504528143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 504657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 504757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 504857de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 504957de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 505057de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 505157de7509SStefano Zampini 505257de7509SStefano Zampini /* further checks */ 5053e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5054e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 5055e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 5056e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 5057e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 505857de7509SStefano Zampini if (reuse && *mat_n) { 505970cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 506057de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 506170cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 506228143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 506370cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 506470cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 506570cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 506670cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 506770cf5478SStefano Zampini } 5068e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5069e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 507057de7509SStefano Zampini 5071e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5072e7931f94SStefano Zampini if (!is_sends) { 507328143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 507457de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5075c8587f34SStefano Zampini } else { 5076e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5077e7931f94SStefano Zampini is_sends_internal = is_sends; 5078c8587f34SStefano Zampini } 5079e7931f94SStefano Zampini 5080e7931f94SStefano Zampini /* get comm */ 5081a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5082e7931f94SStefano Zampini 5083e7931f94SStefano Zampini /* compute number of sends */ 5084e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5085e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5086e7931f94SStefano Zampini 5087e7931f94SStefano Zampini /* compute number of receives */ 5088e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5089785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5090e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5091e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5092e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5093e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5094e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5095e7931f94SStefano Zampini 509628143c3dSStefano Zampini /* restrict comm if requested */ 509728143c3dSStefano Zampini subcomm = 0; 509828143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 509928143c3dSStefano Zampini if (restrict_comm) { 5100779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5101779c1cceSStefano Zampini 510228143c3dSStefano Zampini color = 0; 510353a05cb3SStefano Zampini if (restrict_full) { 510453a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 510553a05cb3SStefano Zampini } else { 510653a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 510753a05cb3SStefano Zampini } 5108b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 510928143c3dSStefano Zampini subcommsize = commsize - subcommsize; 511028143c3dSStefano Zampini /* check if reuse has been requested */ 511157de7509SStefano Zampini if (reuse) { 511228143c3dSStefano Zampini if (*mat_n) { 511328143c3dSStefano Zampini PetscMPIInt subcommsize2; 511428143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 511528143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 511628143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 511728143c3dSStefano Zampini } else { 511828143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 511928143c3dSStefano Zampini } 512028143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5121779c1cceSStefano Zampini PetscMPIInt rank; 5122779c1cceSStefano Zampini 5123779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 512428143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 512528143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 512628143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5127306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 512828143c3dSStefano Zampini } 512928143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 513028143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 513128143c3dSStefano Zampini } else { 513228143c3dSStefano Zampini comm_n = comm; 513328143c3dSStefano Zampini } 513428143c3dSStefano Zampini 5135e7931f94SStefano Zampini /* prepare send/receive buffers */ 5136785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5137e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5138785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5139e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 514028143c3dSStefano Zampini if (nis) { 5141854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 514228143c3dSStefano Zampini } 5143e7931f94SStefano Zampini 514428143c3dSStefano Zampini /* Get data from local matrices */ 51456c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5146e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5147e7931f94SStefano Zampini /* 5148e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5149e7931f94SStefano Zampini send_buffer_idxs should contain: 5150e7931f94SStefano Zampini - MatType_PRIVATE type 5151e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5152e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5153e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5154e7931f94SStefano Zampini */ 51556c4ed002SBarry Smith else { 5156e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 51573bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5158854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5159e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5160e7931f94SStefano Zampini send_buffer_idxs[1] = i; 51613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5162e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 51633bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5164e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5165e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5166e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5167e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5168c8587f34SStefano Zampini } 5169c8587f34SStefano Zampini } 5170e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 517128143c3dSStefano Zampini /* additional is (if any) */ 517228143c3dSStefano Zampini if (nis) { 517328143c3dSStefano Zampini PetscMPIInt psum; 517428143c3dSStefano Zampini PetscInt j; 517528143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 517628143c3dSStefano Zampini PetscInt plen; 517728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 517828143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 517928143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 518028143c3dSStefano Zampini } 5181854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 518228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 518328143c3dSStefano Zampini PetscInt plen; 518428143c3dSStefano Zampini const PetscInt *is_array_idxs; 518528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 518628143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 518728143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 518828143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 518928143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 519028143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 519128143c3dSStefano Zampini } 519228143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 519328143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 519428143c3dSStefano Zampini } 519528143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 519628143c3dSStefano Zampini } 519728143c3dSStefano Zampini 5198e7931f94SStefano Zampini buf_size_idxs = 0; 5199e7931f94SStefano Zampini buf_size_vals = 0; 520028143c3dSStefano Zampini buf_size_idxs_is = 0; 5201e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5202e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5203e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 520428143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5205e7931f94SStefano Zampini } 5206785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5207785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 520895ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5209e7931f94SStefano Zampini 5210e7931f94SStefano Zampini /* get new tags for clean communications */ 5211e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5212e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 521328143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5214e7931f94SStefano Zampini 5215e7931f94SStefano Zampini /* allocate for requests */ 5216785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5217785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 521895ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5219785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5220785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 522195ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5222e7931f94SStefano Zampini 5223e7931f94SStefano Zampini /* communications */ 5224e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5225e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 522628143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5227e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5228e7931f94SStefano Zampini source_dest = onodes[i]; 5229e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5230e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5231e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5232e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 523328143c3dSStefano Zampini if (nis) { 523457de7509SStefano Zampini source_dest = onodes_is[i]; 523528143c3dSStefano 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); 523628143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 523728143c3dSStefano Zampini } 5238e7931f94SStefano Zampini } 5239e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5240e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5241e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5242e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 524328143c3dSStefano Zampini if (nis) { 524428143c3dSStefano 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); 524528143c3dSStefano Zampini } 5246e7931f94SStefano Zampini } 5247e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5248e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5249e7931f94SStefano Zampini 5250e7931f94SStefano Zampini /* assemble new l2g map */ 5251e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5252e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 52539d30be91SStefano Zampini new_local_rows = 0; 5254e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 52559d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5256e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5257e7931f94SStefano Zampini } 52589d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5259e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 52609d30be91SStefano Zampini new_local_rows = 0; 5261e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 52629d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 52639d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5264e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5265e7931f94SStefano Zampini } 52669d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 52679d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5268e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5269e7931f94SStefano Zampini 5270e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5271e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5272e7931f94SStefano 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) */ 5273e7931f94SStefano Zampini if (n_recvs) { 527428143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5275e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5276e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5277e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5278e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5279e7931f94SStefano Zampini break; 5280e7931f94SStefano Zampini } 5281e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5282e7931f94SStefano Zampini } 5283e7931f94SStefano Zampini switch (new_local_type_private) { 528428143c3dSStefano Zampini case MATDENSE_PRIVATE: 528528143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5286e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5287e7931f94SStefano Zampini bs = 1; 528828143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 528928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 529028143c3dSStefano Zampini bs = 1; 529128143c3dSStefano Zampini } 5292e7931f94SStefano Zampini break; 5293e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5294e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5295e7931f94SStefano Zampini bs = 1; 5296e7931f94SStefano Zampini break; 5297e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5298e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5299e7931f94SStefano Zampini break; 5300e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5301e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5302e7931f94SStefano Zampini break; 5303e7931f94SStefano Zampini default: 53049d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5305e7931f94SStefano Zampini break; 5306e7931f94SStefano Zampini } 530728143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 530828143c3dSStefano Zampini new_local_type = MATSEQDENSE; 530928143c3dSStefano Zampini bs = 1; 5310e7931f94SStefano Zampini } 5311e7931f94SStefano Zampini 531270cf5478SStefano Zampini /* create MATIS object if needed */ 531357de7509SStefano Zampini if (!reuse) { 5314e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5315e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 531670cf5478SStefano Zampini } else { 531770cf5478SStefano Zampini /* it also destroys the local matrices */ 531857de7509SStefano Zampini if (*mat_n) { 531970cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 532057de7509SStefano Zampini } else { /* this is a fake object */ 532157de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 532257de7509SStefano Zampini } 532370cf5478SStefano Zampini } 532470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5325e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 53269d30be91SStefano Zampini 53279d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 53289d30be91SStefano Zampini 53299d30be91SStefano Zampini /* Global to local map of received indices */ 53309d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 53319d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 53329d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 53339d30be91SStefano Zampini 53349d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 53359d30be91SStefano Zampini buf_size_idxs = 0; 53369d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 53379d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 53389d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 53399d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 53409d30be91SStefano Zampini } 53419d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 53429d30be91SStefano Zampini 53439d30be91SStefano Zampini /* set preallocation */ 53449d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 53459d30be91SStefano Zampini if (!newisdense) { 53469d30be91SStefano Zampini PetscInt *new_local_nnz=0; 53479d30be91SStefano Zampini 53489d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 53499d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 53509d30be91SStefano Zampini if (n_recvs) { 53519d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 53529d30be91SStefano Zampini } 53539d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 53549d30be91SStefano Zampini PetscInt j; 53559d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 53569d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 53579d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 53589d30be91SStefano Zampini } 53599d30be91SStefano Zampini } else { 53609d30be91SStefano Zampini /* TODO */ 53619d30be91SStefano Zampini } 53629d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 53639d30be91SStefano Zampini } 53649d30be91SStefano Zampini if (new_local_nnz) { 53659d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 53669d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 53679d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 53689d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 53699d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 53709d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 53719d30be91SStefano Zampini } else { 53729d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 53739d30be91SStefano Zampini } 53749d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 53759d30be91SStefano Zampini } else { 53769d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 53779d30be91SStefano Zampini } 5378e7931f94SStefano Zampini 5379e7931f94SStefano Zampini /* set values */ 5380e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 53819d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5382e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5383e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5384e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 53859d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5386e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5387e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5388e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 538928143c3dSStefano Zampini } else { 539028143c3dSStefano Zampini /* TODO */ 5391e7931f94SStefano Zampini } 5392e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5393e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5394e7931f94SStefano Zampini } 5395e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5396e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 539770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 539870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53999d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 54009d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5401e7931f94SStefano Zampini 5402dfd14d43SStefano Zampini #if 0 540328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5404e7931f94SStefano Zampini Vec lvec,rvec; 5405e7931f94SStefano Zampini PetscReal infty_error; 5406e7931f94SStefano Zampini 54072a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5408e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5409e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5410e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 541170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5412e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5413e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5414e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5415e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5416e7931f94SStefano Zampini } 541728143c3dSStefano Zampini #endif 5418e7931f94SStefano Zampini 541928143c3dSStefano Zampini /* assemble new additional is (if any) */ 542028143c3dSStefano Zampini if (nis) { 542128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 542228143c3dSStefano Zampini 542328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5424854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 542528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 542628143c3dSStefano Zampini psum = 0; 542728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 542828143c3dSStefano Zampini for (j=0;j<nis;j++) { 542928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 543028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 543128143c3dSStefano Zampini psum += plen; 543228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 543328143c3dSStefano Zampini } 543428143c3dSStefano Zampini } 5435854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5436854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 543728143c3dSStefano Zampini for (i=1;i<nis;i++) { 543828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 543928143c3dSStefano Zampini } 544028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 544128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 544228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 544328143c3dSStefano Zampini for (j=0;j<nis;j++) { 544428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 544528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 544628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 544728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 544828143c3dSStefano Zampini } 544928143c3dSStefano Zampini } 545028143c3dSStefano Zampini for (i=0;i<nis;i++) { 545128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 545228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 545328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 545428143c3dSStefano Zampini } 545528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 545628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 545728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 545828143c3dSStefano Zampini } 5459e7931f94SStefano Zampini /* free workspace */ 546028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5461e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5462e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5463e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5464e7931f94SStefano Zampini if (isdense) { 5465e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5466e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5467e7931f94SStefano Zampini } else { 5468e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5469e7931f94SStefano Zampini } 547028143c3dSStefano Zampini if (nis) { 547128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 547228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 547328143c3dSStefano Zampini } 5474e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5475e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 547628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5477e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5478e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 547928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5480e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5481e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5482e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5483e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5484e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 548528143c3dSStefano Zampini if (nis) { 548628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 548728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 548828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 548928143c3dSStefano Zampini } 549028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 549128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 549228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 549328143c3dSStefano Zampini for (i=0;i<nis;i++) { 549428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 549528143c3dSStefano Zampini } 549653a05cb3SStefano Zampini *mat_n = NULL; 549728143c3dSStefano Zampini } 5498e7931f94SStefano Zampini PetscFunctionReturn(0); 5499e7931f94SStefano Zampini } 5500a57a6d2fSStefano Zampini 550112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5502af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 550312edc857SStefano Zampini 5504c8587f34SStefano Zampini #undef __FUNCT__ 5505c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5506c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5507c8587f34SStefano Zampini { 5508c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5509c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 551020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 55119881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 551220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 55136e683305SStefano Zampini IS coarse_is,*isarray; 55146e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 551530368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5516f9eb5b7dSStefano Zampini PC pc_temp; 5517c8587f34SStefano Zampini PCType coarse_pc_type; 5518c8587f34SStefano Zampini KSPType coarse_ksp_type; 5519f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 55204f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 55216e683305SStefano Zampini Mat t_coarse_mat_is; 552257de7509SStefano Zampini PetscInt ncoarse; 552368457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 552422bc73bbSStefano Zampini PetscScalar *array; 552557de7509SStefano Zampini MatReuse coarse_mat_reuse; 552657de7509SStefano Zampini PetscBool restr, full_restr, have_void; 55279881197aSStefano Zampini PetscErrorCode ierr; 5528fdc09c96SStefano Zampini 5529c8587f34SStefano Zampini PetscFunctionBegin; 5530c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 553168457ee5SStefano 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 */ 5532fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 55335a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5534fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5535f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5536f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5537f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5538fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 553951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 554051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5541dc4bcba2SStefano Zampini PC pc; 5542dc4bcba2SStefano Zampini PetscBool isbddc; 5543dc4bcba2SStefano Zampini 5544dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5545dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5546dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5547dc4bcba2SStefano Zampini if (isbddc) { 554863c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 554963c961adSStefano Zampini } else { 5550727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 555163c961adSStefano Zampini } 5552fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5553fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5554fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5555f4ddd8eeSStefano Zampini } 5556fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5557fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5558f4ddd8eeSStefano Zampini } 555970cf5478SStefano Zampini /* reset any subassembling information */ 556057de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 556170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 556257de7509SStefano Zampini } 55636e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5564fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5565f4ddd8eeSStefano Zampini } 556657de7509SStefano Zampini /* assemble coarse matrix */ 556757de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 556857de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 556957de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 557057de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 557118a45a71SStefano Zampini } else { 557257de7509SStefano Zampini coarse_mat = NULL; 557357de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 55746e683305SStefano Zampini } 5575e7931f94SStefano Zampini 5576abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5577abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5578abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5579abbbba34SStefano Zampini 5580abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 558122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 558222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 558322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 558422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5585e176bc59SStefano 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); 55866e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 55876e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55886e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5589abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5590abbbba34SStefano Zampini 559157de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 559257de7509SStefano Zampini im_active = !!(pcis->n); 559357de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 559457de7509SStefano Zampini 559557de7509SStefano Zampini /* determine number of process partecipating to coarse solver and compute subassembling pattern */ 559657de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 559757de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 559857de7509SStefano Zampini coarse_mat_is = NULL; 559957de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 560057de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 560157de7509SStefano Zampini full_restr = PETSC_TRUE; 560257de7509SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(pcbddc->coarse_size,pcbddc->coarse_eqs_per_proc); 560357de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 560457de7509SStefano Zampini if (multilevel_requested) { 560557de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 560657de7509SStefano Zampini restr = PETSC_FALSE; 560757de7509SStefano Zampini full_restr = PETSC_FALSE; 560857de7509SStefano Zampini } else { 560957de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 561057de7509SStefano Zampini restr = PETSC_TRUE; 561157de7509SStefano Zampini full_restr = PETSC_TRUE; 561257de7509SStefano Zampini } 561357de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 561457de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 561557de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 561657de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 561757de7509SStefano Zampini PetscInt psum; 561857de7509SStefano Zampini PetscMPIInt size; 561957de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 562057de7509SStefano Zampini else psum = 0; 562157de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 562257de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 562357de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 562457de7509SStefano Zampini } 562557de7509SStefano Zampini /* determine if we can go multilevel */ 562657de7509SStefano Zampini if (multilevel_requested) { 562757de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 562857de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 562957de7509SStefano Zampini } 563057de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 563157de7509SStefano Zampini 5632e4d548c7SStefano Zampini /* dump subassembling pattern */ 5633e4d548c7SStefano Zampini if (pcbddc->dbg_flag && multilevel_allowed) { 5634e4d548c7SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);CHKERRQ(ierr); 5635e4d548c7SStefano Zampini } 5636e4d548c7SStefano Zampini 56376e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 563827b6a85dSStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal)) { /* protects from unneded computations */ 56396e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 56406e683305SStefano Zampini const PetscInt *idxs; 56416e683305SStefano Zampini ISLocalToGlobalMapping tmap; 56426e683305SStefano Zampini 56436e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 56440be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 56456e683305SStefano Zampini /* allocate space for temporary storage */ 5646854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5647854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 56486e683305SStefano Zampini /* allocate for IS array */ 56496e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 56506e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 565127b6a85dSStefano Zampini nisvert = 0; /* nisvert is not used */ 565230368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5653854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 56546e683305SStefano Zampini /* dofs splitting */ 56556e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 56566e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 56576e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 56586e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 56596e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 56606e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 56616e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 566230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 56636e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 56646e683305SStefano Zampini } 56656e683305SStefano Zampini /* neumann boundaries */ 56666e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 56676e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 56686e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 56696e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 56706e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 56716e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 56726e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 567330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 56746e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 56756e683305SStefano Zampini } 56766e683305SStefano Zampini /* free memory */ 56776e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 56786e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 56796e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 56806e683305SStefano Zampini } else { 56816e683305SStefano Zampini nis = 0; 56826e683305SStefano Zampini nisdofs = 0; 56836e683305SStefano Zampini nisneu = 0; 568430368db7SStefano Zampini nisvert = 0; 56856e683305SStefano Zampini isarray = NULL; 56866e683305SStefano Zampini } 56876e683305SStefano Zampini /* destroy no longer needed map */ 56886e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 56896e683305SStefano Zampini 569057de7509SStefano Zampini /* subassemble */ 569157de7509SStefano Zampini if (multilevel_allowed) { 569257de7509SStefano Zampini PetscBool reuse,reuser; 569357de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 569457de7509SStefano Zampini else reuse = PETSC_FALSE; 569557de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 569657de7509SStefano Zampini if (reuser) { 569757de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray);CHKERRQ(ierr); 569874e2c79eSStefano Zampini } else { 569957de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 570074e2c79eSStefano Zampini } 570174e2c79eSStefano Zampini } else { 570257de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 57036e683305SStefano Zampini } 570457de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 570557de7509SStefano Zampini PetscMPIInt size; 570657de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 570757de7509SStefano Zampini if (!multilevel_allowed) { 570857de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 57096e683305SStefano Zampini } else { 571057de7509SStefano Zampini Mat A; 5711779c1cceSStefano Zampini 571257de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 571357de7509SStefano Zampini if (coarse_mat_is) { 571457de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 571557de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 571657de7509SStefano Zampini coarse_mat = coarse_mat_is; 571757de7509SStefano Zampini } 571857de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 571957de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 572057de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5721779c1cceSStefano Zampini } 5722779c1cceSStefano Zampini } 572357de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 572457de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 57256e683305SStefano Zampini 57266e683305SStefano Zampini /* create local to global scatters for coarse problem */ 572768457ee5SStefano Zampini if (compute_vecs) { 57286e683305SStefano Zampini PetscInt lrows; 57296e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 573057de7509SStefano Zampini if (coarse_mat) { 573157de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 57326e683305SStefano Zampini } else { 57336e683305SStefano Zampini lrows = 0; 57346e683305SStefano Zampini } 57356e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 57366e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 57376e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 57386e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 57396e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 57406e683305SStefano Zampini } 57416e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5742c8587f34SStefano Zampini 5743f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5744f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5745f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5746f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5747f9eb5b7dSStefano Zampini } else { 5748f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5749f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5750c8587f34SStefano Zampini } 5751c8587f34SStefano Zampini 57526e683305SStefano Zampini /* print some info if requested */ 57536e683305SStefano Zampini if (pcbddc->dbg_flag) { 57546e683305SStefano Zampini if (!multilevel_allowed) { 57556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57566e683305SStefano Zampini if (multilevel_requested) { 57576e683305SStefano 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); 57586e683305SStefano Zampini } else if (pcbddc->max_levels) { 57596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 57606e683305SStefano Zampini } 57616e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 57626e683305SStefano Zampini } 57636e683305SStefano Zampini } 57646e683305SStefano Zampini 5765f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 576657de7509SStefano Zampini if (coarse_mat) { 57676a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 57686e683305SStefano Zampini if (pcbddc->dbg_flag) { 576957de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 57706e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57716e683305SStefano Zampini } 5772f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5773312be037SStefano Zampini char prefix[256],str_level[16]; 5774e604994aSStefano Zampini size_t len; 577557de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5776422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5777c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5778f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 577957de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5780c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 57816e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5782c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5783c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5784e604994aSStefano Zampini /* prefix */ 5785e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5786e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5787e604994aSStefano Zampini if (!pcbddc->current_level) { 5788e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5789e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5790c8587f34SStefano Zampini } else { 5791e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5792312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5793312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 579434d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5795312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5796e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5797e604994aSStefano Zampini } 5798e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 57993e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 58003e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 58013e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 58023e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5803f9eb5b7dSStefano Zampini /* allow user customization */ 5804f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 58053e3c6dadSStefano Zampini } 58063e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 580751bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 58083e3c6dadSStefano Zampini if (nisdofs) { 58093e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 58103e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 58113e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 58123e3c6dadSStefano Zampini } 58133e3c6dadSStefano Zampini } 58143e3c6dadSStefano Zampini if (nisneu) { 58153e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 58163e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5817312be037SStefano Zampini } 581830368db7SStefano Zampini if (nisvert) { 581930368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 582030368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 582130368db7SStefano Zampini } 5822f9eb5b7dSStefano Zampini 5823f9eb5b7dSStefano Zampini /* get some info after set from options */ 5824f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5825f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 58264f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 58276e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5828f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5829f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5830f9eb5b7dSStefano Zampini } 583139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 58324f3a063dSStefano Zampini if (isredundant) { 58334f3a063dSStefano Zampini KSP inner_ksp; 58344f3a063dSStefano Zampini PC inner_pc; 58354f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 58364f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 58374f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 58384f3a063dSStefano Zampini } 5839f9eb5b7dSStefano Zampini 584057de7509SStefano Zampini /* parameters which miss an API */ 584157de7509SStefano Zampini if (isbddc) { 5842720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5843720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 584457de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 584527b6a85dSStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null; 584627b6a85dSStefano Zampini if (pcbddc_coarse->benign_saddle_point) { 584727b6a85dSStefano Zampini pcbddc_coarse->benign_compute_nonetflux = PETSC_TRUE; 5848720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 584959e48ca4SStefano Zampini if (pcbddc->adaptive_threshold < 1.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE; 5850720d30f9SStefano Zampini } 5851d4d8cf7bSStefano Zampini } 58529881197aSStefano Zampini 58533301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 58545a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 58553301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 58563301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 58573301b35fSStefano Zampini } 58583301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 58593301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 58603301b35fSStefano Zampini } 58613301b35fSStefano Zampini if (pc->pmat->spd_set) { 58623301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 58633301b35fSStefano Zampini } 586427b6a85dSStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) { 586527b6a85dSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 586627b6a85dSStefano Zampini } 58676e683305SStefano Zampini /* set operators */ 58685f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 58696e683305SStefano Zampini if (pcbddc->dbg_flag) { 58706e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 58716e683305SStefano Zampini } 58726e683305SStefano Zampini } 58736e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5874b1ecc7b1SStefano Zampini #if 0 5875b9b85e73SStefano Zampini { 5876b9b85e73SStefano Zampini PetscViewer viewer; 5877b9b85e73SStefano Zampini char filename[256]; 5878b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5879b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 58806a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5881b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5882f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 5883b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5884b9b85e73SStefano Zampini } 5885b9b85e73SStefano Zampini #endif 5886f9eb5b7dSStefano Zampini 5887b8ffe317SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5888298c0119SStefano Zampini #if 0 5889b8ffe317SStefano Zampini if (pcbddc->NullSpace) { 5890b8ffe317SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 589198a51de6SStefano Zampini } 5892298c0119SStefano Zampini #endif 5893b0f5fe93SStefano Zampini /* hack */ 589498a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 589598a51de6SStefano Zampini Vec crhs,csol; 589604708bb6SStefano Zampini 5897f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5898f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5899f347579bSStefano Zampini if (!csol) { 59002a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5901f9eb5b7dSStefano Zampini } 5902f347579bSStefano Zampini if (!crhs) { 59032a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5904f347579bSStefano Zampini } 5905b0f5fe93SStefano Zampini } 5906b0f5fe93SStefano Zampini 5907b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5908b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5909b0f5fe93SStefano Zampini 5910b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 59114f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 59124f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 59134f1b2e48SStefano Zampini } 5914b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5915b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5916b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5917b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5918b0f5fe93SStefano Zampini if (coarse_mat) { 5919b0f5fe93SStefano Zampini Vec nullv; 5920b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5921b0f5fe93SStefano Zampini PetscInt nl; 5922b0f5fe93SStefano Zampini 5923b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5924b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5925b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5926b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5927b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5928b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5929b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5930b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5931b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5932b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5933b0f5fe93SStefano Zampini } 5934b0f5fe93SStefano Zampini } 5935b0f5fe93SStefano Zampini 5936b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5937b0f5fe93SStefano Zampini PetscBool ispreonly; 5938b0f5fe93SStefano Zampini 5939b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5940b0f5fe93SStefano Zampini PetscBool isnull; 5941b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 594227b6a85dSStefano Zampini if (0) { 594330368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5944b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5945b0f5fe93SStefano Zampini } else { 5946b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5947b0f5fe93SStefano Zampini } 5948b0f5fe93SStefano Zampini } else { 5949b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5950b0f5fe93SStefano Zampini } 5951b0f5fe93SStefano Zampini } 5952b0f5fe93SStefano Zampini /* setup coarse ksp */ 5953b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5954cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5955cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 59566e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5957c8587f34SStefano Zampini KSP check_ksp; 59582b510759SStefano Zampini KSPType check_ksp_type; 5959c8587f34SStefano Zampini PC check_pc; 59606e683305SStefano Zampini Vec check_vec,coarse_vec; 59616a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 59622b510759SStefano Zampini PetscInt its; 59636e683305SStefano Zampini PetscBool compute_eigs; 59646e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 59656e683305SStefano Zampini PetscInt neigs; 59668e185a42SStefano Zampini const char *prefix; 5967c8587f34SStefano Zampini 59682b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 59696e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5970422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 597123ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5972f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 5973e4d548c7SStefano Zampini /* prevent from setup unneeded object */ 5974e4d548c7SStefano Zampini ierr = KSPGetPC(check_ksp,&check_pc);CHKERRQ(ierr); 5975e4d548c7SStefano Zampini ierr = PCSetType(check_pc,PCNONE);CHKERRQ(ierr); 59762b510759SStefano Zampini if (ispreonly) { 59772b510759SStefano Zampini check_ksp_type = KSPPREONLY; 59786e683305SStefano Zampini compute_eigs = PETSC_FALSE; 59792b510759SStefano Zampini } else { 5980cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 59816e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5982c8587f34SStefano Zampini } 5983c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 59846e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 59856e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 59866e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5987a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5988a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5989a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5990a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5991c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5992c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5993c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5994c8587f34SStefano Zampini /* create random vec */ 59952701bc32SStefano Zampini ierr = MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);CHKERRQ(ierr); 5996c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5997c8587f34SStefano Zampini if (CoarseNullSpace) { 5998c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5999c8587f34SStefano Zampini } 60006e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 6001c8587f34SStefano Zampini /* solve coarse problem */ 60026e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 6003c8587f34SStefano Zampini if (CoarseNullSpace) { 60046e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 6005c8587f34SStefano Zampini } 6006cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 60076e683305SStefano Zampini if (compute_eigs) { 6008854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 6009854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 60106e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 60116e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 60126e683305SStefano Zampini lambda_min = eigs_r[0]; 60136e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 60142701bc32SStefano Zampini if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */ 60152701bc32SStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);CHKERRQ(ierr); 6016cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6017cbcc2c2aSStefano Zampini } 6018c8587f34SStefano Zampini } 6019c8587f34SStefano Zampini } 6020cbcc2c2aSStefano Zampini 6021c8587f34SStefano Zampini /* check coarse problem residual error */ 60226e683305SStefano Zampini if (pcbddc->dbg_flag) { 60236e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 60246e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 60256e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6026c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 60276e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 60286e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6029779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 60306e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 60316e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 60326e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 60336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6034b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6035b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6036b0f5fe93SStefano Zampini } 60376e683305SStefano Zampini if (compute_eigs) { 60386e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6039deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6040c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 60416e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 60426e683305SStefano 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); 60436e683305SStefano Zampini for (i=0;i<neigs;i++) { 60446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6045c8587f34SStefano Zampini } 60466e683305SStefano Zampini } 60476e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 60486e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 60496e683305SStefano Zampini } 6050e4d548c7SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 60512701bc32SStefano Zampini ierr = VecDestroy(&coarse_vec);CHKERRQ(ierr); 6052c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 60536e683305SStefano Zampini if (compute_eigs) { 60546e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 60556e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6056c8587f34SStefano Zampini } 60576e683305SStefano Zampini } 60586e683305SStefano Zampini } 6059cbcc2c2aSStefano Zampini /* print additional info */ 6060cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 60616e683305SStefano Zampini /* waits until all processes reaches this point */ 60626e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6063cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6064cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6065cbcc2c2aSStefano Zampini } 6066cbcc2c2aSStefano Zampini 60672b510759SStefano Zampini /* free memory */ 6068c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6069fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6070c8587f34SStefano Zampini PetscFunctionReturn(0); 6071c8587f34SStefano Zampini } 6072674ae819SStefano Zampini 6073f34684f1SStefano Zampini #undef __FUNCT__ 6074f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6075f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6076f34684f1SStefano Zampini { 6077f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6078f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6079f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6080dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6081dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 608273be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6083dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6084f34684f1SStefano Zampini PetscErrorCode ierr; 6085f34684f1SStefano Zampini 6086f34684f1SStefano Zampini PetscFunctionBegin; 6087f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 60886c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6089dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 60903bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6091dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6092dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6093dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6094dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6095dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6096dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60976c4ed002SBarry 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); 6098dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6099dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6100dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6101dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6102dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6103f34684f1SStefano Zampini 6104f34684f1SStefano Zampini /* check numbering */ 6105f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6106019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6107dc456d91SStefano Zampini PetscInt i; 6108b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6109f34684f1SStefano Zampini 6110f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6111f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6112f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 61131575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6114019a44ceSStefano Zampini /* counter */ 6115019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6116019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6117019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6118019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6119019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6120019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6121f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6122f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6123727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6124f34684f1SStefano Zampini } 6125f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6126f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6127f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6128e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6129e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6130e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6131e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6132f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6133019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6134f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6135019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 61362c66d082SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi; 613775c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6138b9b85e73SStefano Zampini set_error = PETSC_TRUE; 61392c66d082SStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);CHKERRQ(ierr); 61402c66d082SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d (gid %d) owned by %d processes instead of %d!\n",PetscGlobalRank,i,gi,owned,neigh);CHKERRQ(ierr); 6141f34684f1SStefano Zampini } 6142f34684f1SStefano Zampini } 6143019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6144b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6145f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6146f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6147f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6148f34684f1SStefano Zampini } 6149f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6150f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6151e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6152e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6153f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6154f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6155b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6156ca8b9ea9SStefano Zampini PetscInt *gidxs; 6157ca8b9ea9SStefano Zampini 6158ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 61593bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6160f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6161f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6162f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6163f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 61644bc2dc4bSStefano 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); 6165f34684f1SStefano Zampini } 6166f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6167ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6168f34684f1SStefano Zampini } 6169f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 61701575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6171302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6172f34684f1SStefano Zampini } 61738bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6174f34684f1SStefano Zampini /* get back data */ 6175f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6176f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6177674ae819SStefano Zampini PetscFunctionReturn(0); 6178674ae819SStefano Zampini } 6179674ae819SStefano Zampini 6180e456f2a8SStefano Zampini #undef __FUNCT__ 6181e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6182a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6183e456f2a8SStefano Zampini { 6184e456f2a8SStefano Zampini IS localis_t; 6185a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6186e456f2a8SStefano Zampini PetscScalar *vals; 6187e456f2a8SStefano Zampini PetscErrorCode ierr; 6188e456f2a8SStefano Zampini 6189e456f2a8SStefano Zampini PetscFunctionBegin; 6190a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6191e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6192854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6193e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6194e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6195a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6196a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61971035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6198a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61991035eff8SStefano Zampini } 6200a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6201e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6202e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6203a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6204a7dc3881SStefano Zampini /* now compute set in local ordering */ 6205a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6206a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6207a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6208a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6209a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6210ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6211e456f2a8SStefano Zampini lsize++; 6212e456f2a8SStefano Zampini } 6213e456f2a8SStefano Zampini } 6214854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6215a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6216ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6217e456f2a8SStefano Zampini idxs[lsize++] = i; 6218e456f2a8SStefano Zampini } 6219e456f2a8SStefano Zampini } 6220a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6221a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6222e456f2a8SStefano Zampini *localis = localis_t; 6223e456f2a8SStefano Zampini PetscFunctionReturn(0); 6224e456f2a8SStefano Zampini } 6225906d46d4SStefano Zampini 6226b96c3477SStefano Zampini #undef __FUNCT__ 6227b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 622808122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6229b96c3477SStefano Zampini { 6230a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6231b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6232b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6233a64f4aa4SStefano Zampini Mat S_j; 6234b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6235b96c3477SStefano Zampini PetscBool free_used_adj; 6236b96c3477SStefano Zampini PetscErrorCode ierr; 6237b96c3477SStefano Zampini 6238b96c3477SStefano Zampini PetscFunctionBegin; 6239b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6240b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 624108122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6242b96c3477SStefano Zampini used_xadj = NULL; 6243b96c3477SStefano Zampini used_adjncy = NULL; 6244b96c3477SStefano Zampini } else { 624508122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 624608122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 624708122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 624808122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6249b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6250b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6251b96c3477SStefano Zampini } else { 62522fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6253b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6254b96c3477SStefano Zampini PetscInt nvtxs; 6255b96c3477SStefano Zampini 62562fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 62572fffb893SStefano Zampini if (flg_row) { 6258b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6259b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6260b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6261b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62622fffb893SStefano Zampini } else { 62632fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62642fffb893SStefano Zampini used_xadj = NULL; 62652fffb893SStefano Zampini used_adjncy = NULL; 62662fffb893SStefano Zampini } 62672fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6268b96c3477SStefano Zampini } 6269b96c3477SStefano Zampini } 6270d5574798SStefano Zampini 6271d5574798SStefano Zampini /* setup sub_schurs data */ 6272a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6273df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6274df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6275a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 627691af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,NULL,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 6277a64f4aa4SStefano Zampini } else { 62786816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 6279b7ab4a40SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE;; 6280a3df083aSStefano Zampini PetscInt benign_n; 628172b8c272SStefano Zampini Mat change = NULL; 62829d54b7f4SStefano Zampini Vec scaling = NULL; 628372b8c272SStefano Zampini IS change_primal = NULL; 6284a3df083aSStefano Zampini 62855feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62865feab87aSStefano Zampini PetscInt n_vertices; 62875feab87aSStefano Zampini 62885feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62892034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62905feab87aSStefano Zampini } 629104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 629204708bb6SStefano Zampini if (!isseqaij) { 629304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 629404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 629504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 629604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 629704708bb6SStefano Zampini } else { 6298511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 629904708bb6SStefano Zampini } 630004708bb6SStefano Zampini } 6301a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6302a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6303ca92afb2SStefano Zampini } else { 6304a3df083aSStefano Zampini benign_n = 0; 6305ca92afb2SStefano Zampini } 6306b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6307b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6308b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 630972b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 631022db5ddcSStefano Zampini PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change); 6311b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 631222db5ddcSStefano Zampini need_change = (PetscBool)(!need_change); 6313b7ab4a40SStefano Zampini } 6314b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6315b7ab4a40SStefano 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 */ 6316b7ab4a40SStefano Zampini if (need_change) { 631788c03ad3SStefano Zampini PC_IS *pcisf; 631888c03ad3SStefano Zampini PC_BDDC *pcbddcf; 631988c03ad3SStefano Zampini PC pcf; 632088c03ad3SStefano Zampini 6321e4d548c7SStefano Zampini if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 632288c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 632388c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 632488c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 632588c03ad3SStefano Zampini /* hacks */ 632688c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 632772b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 632872b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 632972b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 633072b8c272SStefano Zampini pcisf->n = pcis->n; 633172b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 633288c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 633388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 633488c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 633588c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 633688c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 633788c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 633872b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 633988c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 634088c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 634172b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 634272b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 634372b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 634472b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 634572b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 634688c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 634772b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 634888c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 634988c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 635088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 635188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 635288c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 635388c03ad3SStefano Zampini pcf->ops->destroy = NULL; 635488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 635588c03ad3SStefano Zampini } 63569d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 635791af6908SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 635872b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 635972b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6360ca92afb2SStefano Zampini } 6361d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6362b96c3477SStefano Zampini 6363b96c3477SStefano Zampini /* free adjacency */ 6364b96c3477SStefano Zampini if (free_used_adj) { 6365b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6366b96c3477SStefano Zampini } 6367b96c3477SStefano Zampini PetscFunctionReturn(0); 6368b96c3477SStefano Zampini } 6369b96c3477SStefano Zampini 6370b96c3477SStefano Zampini #undef __FUNCT__ 6371b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 637208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6373b96c3477SStefano Zampini { 6374b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6375b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6376b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6377b96c3477SStefano Zampini PCBDDCGraph graph; 6378b96c3477SStefano Zampini PetscErrorCode ierr; 6379b96c3477SStefano Zampini 6380b96c3477SStefano Zampini PetscFunctionBegin; 6381b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 638208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63833301b35fSStefano Zampini IS verticesIS,verticescomm; 63843301b35fSStefano Zampini PetscInt vsize,*idxs; 6385b96c3477SStefano Zampini 6386b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63873301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63883301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63893301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63903301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63913301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6392b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63937fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6394441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63953301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6396b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6397b96c3477SStefano Zampini } else { 6398b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6399b96c3477SStefano Zampini } 6400e4d548c7SStefano Zampini /* print some info */ 6401e4d548c7SStefano Zampini if (pcbddc->dbg_flag) { 6402e4d548c7SStefano Zampini IS vertices; 6403e4d548c7SStefano Zampini PetscInt nv,nedges,nfaces; 6404e4d548c7SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6405e4d548c7SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 6406e4d548c7SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 6407e4d548c7SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 6408e4d548c7SStefano Zampini ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6409e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 6410e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 6411e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 6412e4d548c7SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 6413e4d548c7SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6414e4d548c7SStefano Zampini ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6415e4d548c7SStefano Zampini } 6416b96c3477SStefano Zampini 6417b96c3477SStefano Zampini /* sub_schurs init */ 6418a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6419a64f4aa4SStefano Zampini 6420b96c3477SStefano Zampini /* free graph struct */ 642108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6422b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6423b96c3477SStefano Zampini } 6424b96c3477SStefano Zampini PetscFunctionReturn(0); 6425b96c3477SStefano Zampini } 6426fa34dd3eSStefano Zampini 6427fa34dd3eSStefano Zampini #undef __FUNCT__ 6428fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6429fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6430fa34dd3eSStefano Zampini { 6431fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6432fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6433fa34dd3eSStefano Zampini PetscErrorCode ierr; 6434fa34dd3eSStefano Zampini 6435fa34dd3eSStefano Zampini PetscFunctionBegin; 6436fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6437fa34dd3eSStefano Zampini IS zerodiag = NULL; 64384f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6439fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 64404f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 644175c01103SStefano Zampini PetscReal norm; 6442fa34dd3eSStefano Zampini PetscInt i; 6443fa34dd3eSStefano Zampini 6444fa34dd3eSStefano Zampini /* B0 and B0_B */ 6445fa34dd3eSStefano Zampini if (zerodiag) { 6446fa34dd3eSStefano Zampini IS dummy; 6447fa34dd3eSStefano Zampini 64484f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 64494f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6450fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6451fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6452fa34dd3eSStefano Zampini } 6453fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6454fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6455fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6456fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6457fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6458fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6459fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6460fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6461fa34dd3eSStefano Zampini /* S_j */ 6462fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6463fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6464fa34dd3eSStefano Zampini 6465fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6466fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6467fa34dd3eSStefano Zampini /* continuous in primal space */ 6468fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6469fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6470fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6471fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 64724f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 64734f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6474fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6475fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6476fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6477fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6478fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6479fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6480fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6481fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6482fa34dd3eSStefano Zampini 6483fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6484fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6485fa34dd3eSStefano Zampini /* local with Schur */ 6486fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6487fa34dd3eSStefano Zampini if (zerodiag) { 6488fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64894f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6490fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6491fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6492fa34dd3eSStefano Zampini } 6493fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6494fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6495fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6496fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6497fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6498fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6499fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6500fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6501fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6502fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6503fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6504fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6505fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6506fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6507fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6508fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6509fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6510fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6511fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6512fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6513fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6514fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6515fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6516fa34dd3eSStefano Zampini if (zerodiag) { 6517fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6518fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 65194f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6520fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6521fa34dd3eSStefano Zampini } 6522fa34dd3eSStefano Zampini /* BDDC */ 6523fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6524fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6525fa34dd3eSStefano Zampini 6526fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6527fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6528fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6529fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 65304f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 65314f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6532fa34dd3eSStefano Zampini } 65334f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6534fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6535fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6536fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6537fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6538fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6539fa34dd3eSStefano Zampini } 6540fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6541fa34dd3eSStefano Zampini } 6542