11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5674ae819SStefano Zampini 6a3df083aSStefano Zampini #undef __FUNCT__ 7a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 8a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 9a3df083aSStefano Zampini { 10a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 11a3df083aSStefano Zampini PetscErrorCode ierr; 12a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 13a3df083aSStefano Zampini 14a3df083aSStefano Zampini PetscFunctionBegin; 15a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 16a3df083aSStefano Zampini if (transpose) { 17a3df083aSStefano Zampini apply_right = ctx->apply_left; 18a3df083aSStefano Zampini apply_left = ctx->apply_right; 19a3df083aSStefano Zampini } else { 20a3df083aSStefano Zampini apply_right = ctx->apply_right; 21a3df083aSStefano Zampini apply_left = ctx->apply_left; 22a3df083aSStefano Zampini } 23a3df083aSStefano Zampini reset_x = PETSC_FALSE; 24a3df083aSStefano Zampini if (apply_right) { 25a3df083aSStefano Zampini const PetscScalar *ax; 26a3df083aSStefano Zampini PetscInt nl,i; 27a3df083aSStefano Zampini 28a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 29a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 30a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 31a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 32a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 33a3df083aSStefano Zampini PetscScalar sum,val; 34a3df083aSStefano Zampini const PetscInt *idxs; 35a3df083aSStefano Zampini PetscInt nz,j; 36a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 37a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 38a3df083aSStefano Zampini sum = 0.; 39a3df083aSStefano Zampini if (ctx->apply_p0) { 40a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 41a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 42a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 43a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 44a3df083aSStefano Zampini } 45a3df083aSStefano Zampini } else { 46a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 47a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 48a3df083aSStefano Zampini } 49a3df083aSStefano Zampini } 50a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 51a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 52a3df083aSStefano Zampini } 53a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 54a3df083aSStefano Zampini reset_x = PETSC_TRUE; 55a3df083aSStefano Zampini } 56a3df083aSStefano Zampini if (transpose) { 57a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 58a3df083aSStefano Zampini } else { 59a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 60a3df083aSStefano Zampini } 61a3df083aSStefano Zampini if (reset_x) { 62a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 63a3df083aSStefano Zampini } 64a3df083aSStefano Zampini if (apply_left) { 65a3df083aSStefano Zampini PetscScalar *ay; 66a3df083aSStefano Zampini PetscInt i; 67a3df083aSStefano Zampini 68a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 69a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 70a3df083aSStefano Zampini PetscScalar sum,val; 71a3df083aSStefano Zampini const PetscInt *idxs; 72a3df083aSStefano Zampini PetscInt nz,j; 73a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 74a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 75a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 76a3df083aSStefano Zampini if (ctx->apply_p0) { 77a3df083aSStefano Zampini sum = 0.; 78a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 79a3df083aSStefano Zampini sum += ay[idxs[j]]; 80a3df083aSStefano Zampini ay[idxs[j]] += val; 81a3df083aSStefano Zampini } 82a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 83a3df083aSStefano Zampini } else { 84a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 85a3df083aSStefano Zampini ay[idxs[j]] += val; 86a3df083aSStefano Zampini } 87a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 88a3df083aSStefano Zampini } 89a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 90a3df083aSStefano Zampini } 91a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 92a3df083aSStefano Zampini } 93a3df083aSStefano Zampini PetscFunctionReturn(0); 94a3df083aSStefano Zampini } 95a3df083aSStefano Zampini 96a3df083aSStefano Zampini #undef __FUNCT__ 97a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 98a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 99a3df083aSStefano Zampini { 100a3df083aSStefano Zampini PetscErrorCode ierr; 101a3df083aSStefano Zampini 102a3df083aSStefano Zampini PetscFunctionBegin; 103a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 104a3df083aSStefano Zampini PetscFunctionReturn(0); 105a3df083aSStefano Zampini } 106a3df083aSStefano Zampini 107a3df083aSStefano Zampini #undef __FUNCT__ 108a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 109a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 110a3df083aSStefano Zampini { 111a3df083aSStefano Zampini PetscErrorCode ierr; 112a3df083aSStefano Zampini 113a3df083aSStefano Zampini PetscFunctionBegin; 114a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 115a3df083aSStefano Zampini PetscFunctionReturn(0); 116a3df083aSStefano Zampini } 117a3df083aSStefano Zampini 118a3df083aSStefano Zampini #undef __FUNCT__ 119a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 120a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 121a3df083aSStefano Zampini { 122a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 123a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 124a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 125a3df083aSStefano Zampini PetscErrorCode ierr; 126a3df083aSStefano Zampini 127a3df083aSStefano Zampini PetscFunctionBegin; 128a3df083aSStefano Zampini if (!restore) { 1291dd7afcfSStefano Zampini Mat A_IB,A_BI; 130a3df083aSStefano Zampini PetscScalar *work; 131df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 132a3df083aSStefano Zampini 1331dd7afcfSStefano Zampini if (pcbddc->benign_original_mat) { 1341dd7afcfSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 1351dd7afcfSStefano Zampini } 1361dd7afcfSStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) { 1371dd7afcfSStefano Zampini PetscFunctionReturn(0); 1381dd7afcfSStefano Zampini } 139a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 140a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 141a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 142a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 145a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 146a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 147a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 148a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 149a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 150a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 151059032f7SStefano Zampini if (reuse) { 152a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 1531dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 154059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 155059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 156059032f7SStefano Zampini PetscInt i; 157059032f7SStefano Zampini 158059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 159059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 160059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 161059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 162059032f7SStefano Zampini } 163059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 1641dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 165059032f7SStefano Zampini } 166a3df083aSStefano Zampini ctx->A = pcis->A_IB; 167a3df083aSStefano Zampini ctx->work = work; 168a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 169a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 170a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 171a3df083aSStefano Zampini pcis->A_IB = A_IB; 172a3df083aSStefano Zampini 173a3df083aSStefano Zampini /* A_BI as A_IB^T */ 174a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 175a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 176a3df083aSStefano Zampini pcis->A_BI = A_BI; 177a3df083aSStefano Zampini } else { 1781dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 1791dd7afcfSStefano Zampini PetscFunctionReturn(0); 1801dd7afcfSStefano Zampini } 181a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 182a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 183a3df083aSStefano Zampini pcis->A_IB = ctx->A; 1841dd7afcfSStefano Zampini ctx->A = NULL; 1851dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 1861dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 1871dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 1881dd7afcfSStefano Zampini if (ctx->free) { 189059032f7SStefano Zampini PetscInt i; 1901dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 191059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 192059032f7SStefano Zampini } 193059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 194059032f7SStefano Zampini } 195a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 196a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 197a3df083aSStefano Zampini } 198a3df083aSStefano Zampini PetscFunctionReturn(0); 199a3df083aSStefano Zampini } 200a3df083aSStefano Zampini 201a3df083aSStefano Zampini /* used just in bddc debug mode */ 202a3df083aSStefano Zampini #undef __FUNCT__ 203a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 204a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 205a3df083aSStefano Zampini { 206a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 207a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 208a3df083aSStefano Zampini Mat An; 209a3df083aSStefano Zampini PetscErrorCode ierr; 210a3df083aSStefano Zampini 211a3df083aSStefano Zampini PetscFunctionBegin; 212a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 213a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 214a3df083aSStefano Zampini if (is1) { 215a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 216a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 217a3df083aSStefano Zampini } else { 218a3df083aSStefano Zampini *B = An; 219a3df083aSStefano Zampini } 220a3df083aSStefano Zampini PetscFunctionReturn(0); 221a3df083aSStefano Zampini } 222a3df083aSStefano Zampini 2231cf9b237SStefano Zampini /* TODO: add reuse flag */ 2241cf9b237SStefano Zampini #undef __FUNCT__ 2251cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2261cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2271cf9b237SStefano Zampini { 2281cf9b237SStefano Zampini Mat Bt; 2291cf9b237SStefano Zampini PetscScalar *a,*bdata; 2301cf9b237SStefano Zampini const PetscInt *ii,*ij; 2311cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2321cf9b237SStefano Zampini PetscBool flg_row; 2331cf9b237SStefano Zampini PetscErrorCode ierr; 2341cf9b237SStefano Zampini 2351cf9b237SStefano Zampini PetscFunctionBegin; 2361cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2371cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2381cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2391cf9b237SStefano Zampini nnz = n; 2401cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2411cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2421cf9b237SStefano Zampini } 2431cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2441cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2451cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2461cf9b237SStefano Zampini nnz = 0; 2471cf9b237SStefano Zampini bii[0] = 0; 2481cf9b237SStefano Zampini for (i=0;i<n;i++) { 2491cf9b237SStefano Zampini PetscInt j; 2501cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2511cf9b237SStefano Zampini PetscScalar entry = a[j]; 2521cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2531cf9b237SStefano Zampini bij[nnz] = ij[j]; 2541cf9b237SStefano Zampini bdata[nnz] = entry; 2551cf9b237SStefano Zampini nnz++; 2561cf9b237SStefano Zampini } 2571cf9b237SStefano Zampini } 2581cf9b237SStefano Zampini bii[i+1] = nnz; 2591cf9b237SStefano Zampini } 2601cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2611cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2621cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2631cf9b237SStefano Zampini { 2641cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2651cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2661cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2671cf9b237SStefano Zampini } 2681cf9b237SStefano Zampini *B = Bt; 2691cf9b237SStefano Zampini PetscFunctionReturn(0); 2701cf9b237SStefano Zampini } 2711cf9b237SStefano Zampini 272674ae819SStefano Zampini #undef __FUNCT__ 2734f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2744f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2754f1b2e48SStefano Zampini { 2764f1b2e48SStefano Zampini Mat B; 2774f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2784f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2794f1b2e48SStefano Zampini PCBDDCGraph graph; 2804f1b2e48SStefano Zampini PetscInt i,n; 2814f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 2824f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 2834f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 2844f1b2e48SStefano Zampini PetscErrorCode ierr; 2854f1b2e48SStefano Zampini 2864f1b2e48SStefano Zampini PetscFunctionBegin; 28763c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 28863c961adSStefano Zampini *ncc = 0; 28963c961adSStefano Zampini *cc = NULL; 29063c961adSStefano Zampini PetscFunctionReturn(0); 29163c961adSStefano Zampini } 2924f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 2934f1b2e48SStefano Zampini if (!isseqaij && filter) { 2941cf9b237SStefano Zampini PetscBool isseqdense; 2951cf9b237SStefano Zampini 2961cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 2971cf9b237SStefano Zampini if (!isseqdense) { 2984f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 2991cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3001cf9b237SStefano Zampini PetscScalar *array; 3011cf9b237SStefano Zampini PetscReal chop=1.e-6; 3021cf9b237SStefano Zampini 3031cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3041cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3051cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3061cf9b237SStefano Zampini for (i=0;i<n;i++) { 3071cf9b237SStefano Zampini PetscInt j; 3081cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3091cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3101cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3111cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3121cf9b237SStefano Zampini } 3131cf9b237SStefano Zampini } 3141cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3159d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 3161cf9b237SStefano Zampini } 3174f1b2e48SStefano Zampini } else { 3184f1b2e48SStefano Zampini B = A; 3194f1b2e48SStefano Zampini } 3204f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3214f1b2e48SStefano Zampini 3224f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3234f1b2e48SStefano Zampini if (filter) { 3244f1b2e48SStefano Zampini PetscScalar *data; 3254f1b2e48SStefano Zampini PetscInt j,cum; 3264f1b2e48SStefano Zampini 3274f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3284f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3294f1b2e48SStefano Zampini cum = 0; 3304f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3314f1b2e48SStefano Zampini PetscInt t; 3324f1b2e48SStefano Zampini 3334f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3344f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3354f1b2e48SStefano Zampini continue; 3364f1b2e48SStefano Zampini } 3374f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini t = xadj_filtered[i]; 3404f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3414f1b2e48SStefano Zampini cum += t; 3424f1b2e48SStefano Zampini } 3434f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3444f1b2e48SStefano Zampini } else { 3454f1b2e48SStefano Zampini xadj_filtered = NULL; 3464f1b2e48SStefano Zampini adjncy_filtered = NULL; 3474f1b2e48SStefano Zampini } 3484f1b2e48SStefano Zampini 3494f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3504f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3514f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3524f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3534f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3544f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3554f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3564f1b2e48SStefano Zampini if (xadj_filtered) { 3574f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3584f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3594f1b2e48SStefano Zampini } else { 3604f1b2e48SStefano Zampini graph->xadj = xadj; 3614f1b2e48SStefano Zampini graph->adjncy = adjncy; 3624f1b2e48SStefano Zampini } 3634f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3644f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3654f1b2e48SStefano Zampini /* partial clean up */ 3664f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3674f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3681cf9b237SStefano Zampini if (A != B) { 3694f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3704f1b2e48SStefano Zampini } 3714f1b2e48SStefano Zampini 3724f1b2e48SStefano Zampini /* get back data */ 3731cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3741cf9b237SStefano Zampini if (cc) { 3754f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3764f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3774f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 3784f1b2e48SStefano Zampini } 3794f1b2e48SStefano Zampini *cc = cc_n; 3801cf9b237SStefano Zampini } 3814f1b2e48SStefano Zampini /* clean up graph */ 3824f1b2e48SStefano Zampini graph->xadj = 0; 3834f1b2e48SStefano Zampini graph->adjncy = 0; 3844f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 3854f1b2e48SStefano Zampini PetscFunctionReturn(0); 3864f1b2e48SStefano Zampini } 3874f1b2e48SStefano Zampini 3884f1b2e48SStefano Zampini #undef __FUNCT__ 3895408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 3905408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 3915408967cSStefano Zampini { 3925408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3935408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 394dee84bffSStefano Zampini IS dirIS = NULL; 3954f1b2e48SStefano Zampini PetscInt i; 3965408967cSStefano Zampini PetscErrorCode ierr; 3975408967cSStefano Zampini 3985408967cSStefano Zampini PetscFunctionBegin; 399dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4005408967cSStefano Zampini if (zerodiag) { 4015408967cSStefano Zampini Mat A; 4025408967cSStefano Zampini Vec vec3_N; 4035408967cSStefano Zampini PetscScalar *vals; 4045408967cSStefano Zampini const PetscInt *idxs; 405d12d3064SStefano Zampini PetscInt nz,*count; 4065408967cSStefano Zampini 4075408967cSStefano Zampini /* p0 */ 4085408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4095408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4105408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4115408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4124f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4135408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4145408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4155408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4165408967cSStefano Zampini /* v_I */ 4175408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4185408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4195408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4205408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4215408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4225408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4235408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4245408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4255408967cSStefano Zampini if (dirIS) { 4265408967cSStefano Zampini PetscInt n; 4275408967cSStefano Zampini 4285408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4295408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4305408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4315408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4325408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4335408967cSStefano Zampini } 4345408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4355408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4365408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4375408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4385408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 441*fbfcb133SStefano Zampini if (PetscAbsScalar(vals[0]) > 1.e-1) { 442b9b0e38cSStefano 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])); 4435408967cSStefano Zampini } 4445408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4455408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 446d12d3064SStefano Zampini 447d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 448d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 449d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 450d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 451d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 452d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 453d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 454d12d3064SStefano Zampini if (count[idxs[i]]) { 455d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 456d12d3064SStefano Zampini } 457d4d8cf7bSStefano Zampini } 458d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 459d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4605408967cSStefano Zampini } 461dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4625408967cSStefano Zampini 4635408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4645408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4665408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4674f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4685408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4694f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4704f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4714f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 4724f1b2e48SStefano Zampini } 4735408967cSStefano Zampini } 4745408967cSStefano Zampini PetscFunctionReturn(0); 4755408967cSStefano Zampini } 4765408967cSStefano Zampini 4775408967cSStefano Zampini #undef __FUNCT__ 478339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 479339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 480339f8db1SStefano Zampini { 481339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4824f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 483b0f5fe93SStefano Zampini PetscInt nz,n; 4844f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 485339f8db1SStefano Zampini PetscErrorCode ierr; 486339f8db1SStefano Zampini 487339f8db1SStefano Zampini PetscFunctionBegin; 4889f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4899f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 490a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 491a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 492a3df083aSStefano Zampini } 493a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 494a3df083aSStefano Zampini pcbddc->benign_n = 0; 4954f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4964f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 4974f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 4984f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 4994f1b2e48SStefano Zampini since the local Schur complements are SPD 5004f1b2e48SStefano Zampini */ 5014f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5024f1b2e48SStefano Zampini have_null = PETSC_TRUE; 50340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5044f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5054f1b2e48SStefano Zampini 5064f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5074f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5084f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5094f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 510ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 51140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 51240fa8d13SStefano Zampini if (!sorted) { 51340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 51440fa8d13SStefano Zampini } 51540fa8d13SStefano Zampini } else { 51640fa8d13SStefano Zampini pressures = NULL; 51740fa8d13SStefano Zampini } 51897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 51997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 52097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 521339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 522339f8db1SStefano Zampini if (!sorted) { 523339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 524339f8db1SStefano Zampini } 525339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5264f1b2e48SStefano Zampini if (!nz) { 5274f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5284f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 52940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 53040fa8d13SStefano Zampini } 5314f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5324f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5334f1b2e48SStefano Zampini zerodiag_subs = NULL; 5344f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5354f1b2e48SStefano Zampini if (has_null_pressures) { 5364f1b2e48SStefano Zampini IS *subs; 5374f1b2e48SStefano Zampini PetscInt nsubs,i; 5384f1b2e48SStefano Zampini 5394f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5404f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5414f1b2e48SStefano Zampini if (nsubs > 1) { 5424f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5444f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5454f1b2e48SStefano Zampini IS t_zerodiag_subs; 5464f1b2e48SStefano Zampini PetscInt nl; 5474f1b2e48SStefano Zampini 5484f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5494f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5504f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5514f1b2e48SStefano Zampini if (nl) { 5524f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5534f1b2e48SStefano Zampini 5544f1b2e48SStefano Zampini if (pressures) { 5554f1b2e48SStefano Zampini IS t_pressure_subs; 5564f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5574f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5584f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5594f1b2e48SStefano Zampini } 5604f1b2e48SStefano Zampini if (valid) { 5614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini pcbddc->benign_n++; 5634f1b2e48SStefano Zampini } else { 5644f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5654f1b2e48SStefano Zampini } 5664f1b2e48SStefano Zampini } 5674f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini } 5704f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5714f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5724f1b2e48SStefano Zampini if (pressures) { 5734f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5744f1b2e48SStefano Zampini } 5754f1b2e48SStefano Zampini if (valid) { 5764f1b2e48SStefano Zampini pcbddc->benign_n = 1; 577ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5784f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5794f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5804f1b2e48SStefano Zampini } 5814f1b2e48SStefano Zampini } 5824f1b2e48SStefano Zampini } 5834f1b2e48SStefano Zampini 5844f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 585b9b0e38cSStefano Zampini PetscInt n; 586b9b0e38cSStefano Zampini 5874f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 589b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 590b9b0e38cSStefano Zampini if (n) { 5914f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5924f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5934f1b2e48SStefano Zampini } 594b9b0e38cSStefano Zampini } 5954f1b2e48SStefano Zampini 5964f1b2e48SStefano Zampini /* final check for null pressures */ 5974f1b2e48SStefano Zampini if (zerodiag && pressures) { 5984f1b2e48SStefano Zampini PetscInt nz,np; 5994f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6004f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 6014f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 6024f1b2e48SStefano Zampini } 6034f1b2e48SStefano Zampini 6044f1b2e48SStefano Zampini if (recompute_zerodiag) { 6054f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6074f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6084f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6094f1b2e48SStefano Zampini } else { 6104f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6114f1b2e48SStefano Zampini 6124f1b2e48SStefano Zampini nzn = 0; 6134f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6144f1b2e48SStefano Zampini PetscInt ns; 6154f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6164f1b2e48SStefano Zampini nzn += ns; 6174f1b2e48SStefano Zampini } 6184f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6194f1b2e48SStefano Zampini nzn = 0; 6204f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6214f1b2e48SStefano Zampini PetscInt ns,*idxs; 6224f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6234f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6244f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6254f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6264f1b2e48SStefano Zampini nzn += ns; 6274f1b2e48SStefano Zampini } 6284f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6294f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6304f1b2e48SStefano Zampini } 6314f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6324f1b2e48SStefano Zampini } 6334f1b2e48SStefano Zampini 634b3afcdbeSStefano Zampini /* no-net-flux */ 635b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 636b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 637b3afcdbeSStefano Zampini MatNullSpace near_null_space; 638b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 639b3afcdbeSStefano Zampini Vec quad_vec; 640b3afcdbeSStefano Zampini PetscScalar *pvals; 641b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 642b3afcdbeSStefano Zampini 643b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 644b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 645b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 646b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 647b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 648b3afcdbeSStefano Zampini 649b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 650b3afcdbeSStefano Zampini np = 0; 651b3afcdbeSStefano Zampini if (zerodiag) { 652b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 653b3afcdbeSStefano Zampini } else if (pressures) { 654b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 655b3afcdbeSStefano Zampini } 656b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 657b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 658b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 659b3afcdbeSStefano Zampini if (np) { 660b3afcdbeSStefano Zampini const PetscInt *pidxs; 661b3afcdbeSStefano Zampini 662b3afcdbeSStefano Zampini if (zerodiag) { 663b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 664b3afcdbeSStefano Zampini } else if (pressures) { 665b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 666b3afcdbeSStefano Zampini } 667b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 668b3afcdbeSStefano Zampini if (zerodiag) { 669b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 670b3afcdbeSStefano Zampini } else if (pressures) { 671b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 672b3afcdbeSStefano Zampini } 673b3afcdbeSStefano Zampini } 674b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 675b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 676b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 677b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 678b3afcdbeSStefano Zampini 679b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 680b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 681b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = -1; 682b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 683b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 684b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 685b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 686b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 687b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 688b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) { 689b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 690b3afcdbeSStefano Zampini pvals[i] = 0.; 691b3afcdbeSStefano Zampini } 692b3afcdbeSStefano Zampini } 693b3afcdbeSStefano Zampini ierr = PetscMalloc1(matis->sf_nleaves,&dummyins);CHKERRQ(ierr); 694b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) dummyins[i] = i; 695b3afcdbeSStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf_nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 696b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 697b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 698b3afcdbeSStefano Zampini 699b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 700b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 701b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 702b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 703b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 704b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 705b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 706b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 707b3afcdbeSStefano Zampini } 708b3afcdbeSStefano Zampini 709b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7104f1b2e48SStefano Zampini if (has_null_pressures) { 7114f1b2e48SStefano Zampini IS zerodiagc; 7124f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7134f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7144f1b2e48SStefano Zampini 7154f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 716339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 717339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 718339f8db1SStefano Zampini /* local change of basis for pressures */ 719339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 72097d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 721339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 722339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 723339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7244f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7264f1b2e48SStefano Zampini PetscInt nzs,j; 7274f1b2e48SStefano Zampini 7284f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7294f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7304f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7314f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7324f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7334f1b2e48SStefano Zampini } 734339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 735339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 736339f8db1SStefano Zampini /* set identity on velocities */ 737339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 738339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 739339f8db1SStefano Zampini } 7404f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7414f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7429f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7434f1b2e48SStefano 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); 744339f8db1SStefano Zampini /* set change on pressures */ 7454f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7464f1b2e48SStefano Zampini PetscScalar *array; 7474f1b2e48SStefano Zampini PetscInt nzs; 7484f1b2e48SStefano Zampini 7494f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7504f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7514f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 752339f8db1SStefano Zampini PetscScalar vals[2]; 753339f8db1SStefano Zampini PetscInt cols[2]; 754339f8db1SStefano Zampini 755339f8db1SStefano Zampini cols[0] = idxs[i]; 7564f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 757339f8db1SStefano Zampini vals[0] = 1.; 758b0f5fe93SStefano Zampini vals[1] = 1.; 7594f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 760339f8db1SStefano Zampini } 7614f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7624f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7634f1b2e48SStefano Zampini array[nzs-1] = 1.; 7644f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7654f1b2e48SStefano Zampini /* store local idxs for p0 */ 7664f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7674f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 768339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7694f1b2e48SStefano Zampini } 770339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 771339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 772a3df083aSStefano Zampini /* project if needed */ 773a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7741dd7afcfSStefano Zampini Mat M; 7751dd7afcfSStefano Zampini 7761dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 777339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 7781dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 7791dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 780a3df083aSStefano Zampini } 7814f1b2e48SStefano Zampini /* store global idxs for p0 */ 7824f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 783339f8db1SStefano Zampini } 784ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7854f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 786b0f5fe93SStefano Zampini 787b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 788b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 789339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 790339f8db1SStefano Zampini PetscFunctionReturn(0); 791339f8db1SStefano Zampini } 792339f8db1SStefano Zampini 793339f8db1SStefano Zampini #undef __FUNCT__ 794015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 795015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 796efc2fbd9SStefano Zampini { 797efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 798de9d7bd0SStefano Zampini PetscScalar *array; 799efc2fbd9SStefano Zampini PetscErrorCode ierr; 800efc2fbd9SStefano Zampini 801efc2fbd9SStefano Zampini PetscFunctionBegin; 802efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 803efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 8044f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 805efc2fbd9SStefano Zampini } 806de9d7bd0SStefano Zampini if (get) { 807efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8084f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8094f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 810efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 811de9d7bd0SStefano Zampini } else { 812de9d7bd0SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 813de9d7bd0SStefano Zampini ierr = PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 814de9d7bd0SStefano Zampini ierr = PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);CHKERRQ(ierr); 815de9d7bd0SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 816efc2fbd9SStefano Zampini } 817efc2fbd9SStefano Zampini PetscFunctionReturn(0); 818efc2fbd9SStefano Zampini } 819efc2fbd9SStefano Zampini 820efc2fbd9SStefano Zampini #undef __FUNCT__ 821c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 822c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 823c263805aSStefano Zampini { 824c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 825c263805aSStefano Zampini PetscErrorCode ierr; 826c263805aSStefano Zampini 827c263805aSStefano Zampini PetscFunctionBegin; 828c263805aSStefano Zampini /* TODO: add error checking 829c263805aSStefano Zampini - avoid nested pop (or push) calls. 830c263805aSStefano Zampini - cannot push before pop. 8311c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 832c263805aSStefano Zampini */ 8334f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 834efc2fbd9SStefano Zampini PetscFunctionReturn(0); 835efc2fbd9SStefano Zampini } 836c263805aSStefano Zampini if (pop) { 837a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8384f1b2e48SStefano Zampini IS is_p0; 8394f1b2e48SStefano Zampini MatReuse reuse; 840c263805aSStefano Zampini 841c263805aSStefano Zampini /* extract B_0 */ 8424f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8434f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8444f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8454f1b2e48SStefano Zampini } 8464f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8474f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 848c263805aSStefano Zampini /* remove rows and cols from local problem */ 849c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 85097d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8514f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8524f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 853a3df083aSStefano Zampini } else { 854a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 855a3df083aSStefano Zampini PetscScalar *vals; 856a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 857a3df083aSStefano Zampini 858a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 859a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 860a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8610b5adadeSStefano Zampini PetscInt *nnz; 862a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 863a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 864a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 865331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 866331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 867331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 868331e053bSStefano Zampini nnz[i] = n - nnz[i]; 869331e053bSStefano Zampini } 870331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 871331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 872331e053bSStefano Zampini } 873a3df083aSStefano Zampini 874a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 875a3df083aSStefano Zampini PetscScalar *array; 876a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 877a3df083aSStefano Zampini 878a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 879a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 880a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 881a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 882a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 883a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 884a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 885a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 886a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 887a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 888a3df083aSStefano Zampini cum = 0; 889a3df083aSStefano Zampini for (j=0;j<n;j++) { 890a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 891a3df083aSStefano Zampini vals[cum] = array[j]; 892a3df083aSStefano Zampini idxs_ins[cum] = j; 893a3df083aSStefano Zampini cum++; 894a3df083aSStefano Zampini } 895a3df083aSStefano Zampini } 896a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 897a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 898a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 899a3df083aSStefano Zampini } 900a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 901a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 902a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 903a3df083aSStefano Zampini } 904c263805aSStefano Zampini } else { /* push */ 905a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9064f1b2e48SStefano Zampini PetscInt i; 9074f1b2e48SStefano Zampini 9084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9094f1b2e48SStefano Zampini PetscScalar *B0_vals; 9104f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9114f1b2e48SStefano Zampini 9124f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9134f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9147b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9154f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9164f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9174f1b2e48SStefano Zampini } 918c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 919c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 920a3df083aSStefano Zampini } else { 921a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 922a3df083aSStefano Zampini } 923c263805aSStefano Zampini } 924c263805aSStefano Zampini PetscFunctionReturn(0); 925c263805aSStefano Zampini } 926c263805aSStefano Zampini 927c263805aSStefano Zampini #undef __FUNCT__ 928b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 92908122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 930b1b3d7a2SStefano Zampini { 931b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 93208122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 93308122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 93408122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 93508122e43SStefano Zampini PetscScalar *work,lwork; 93608122e43SStefano Zampini PetscScalar *St,*S,*eigv; 93708122e43SStefano Zampini PetscScalar *Sarray,*Starray; 93808122e43SStefano Zampini PetscReal *eigs,thresh; 9391b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 940f6f667cfSStefano Zampini PetscBool allocated_S_St; 94108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 94208122e43SStefano Zampini PetscReal *rwork; 94308122e43SStefano Zampini #endif 944b1b3d7a2SStefano Zampini PetscErrorCode ierr; 945b1b3d7a2SStefano Zampini 946b1b3d7a2SStefano Zampini PetscFunctionBegin; 947af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 948af25d912SStefano 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); 94906a4e24aSStefano Zampini 950fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 951fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 952fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 953fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9541575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 955fd14bc51SStefano Zampini } 956fd14bc51SStefano Zampini 957e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 958e496cd5dSStefano 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); 959e496cd5dSStefano Zampini } 960e496cd5dSStefano Zampini 96108122e43SStefano Zampini /* max size of subsets */ 96208122e43SStefano Zampini mss = 0; 96308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 96408122e43SStefano Zampini PetscInt subset_size; 965862806e4SStefano Zampini 96608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 96708122e43SStefano Zampini mss = PetscMax(mss,subset_size); 96808122e43SStefano Zampini } 96908122e43SStefano Zampini 97008122e43SStefano Zampini /* min/max and threshold */ 97108122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 972f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 97308122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 974f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 975f6f667cfSStefano Zampini if (nmin) { 976f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 977f6f667cfSStefano Zampini } 97808122e43SStefano Zampini 97908122e43SStefano Zampini /* allocate lapack workspace */ 98008122e43SStefano Zampini cum = cum2 = 0; 98108122e43SStefano Zampini maxneigs = 0; 98208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 98308122e43SStefano Zampini PetscInt n,subset_size; 984f6f667cfSStefano Zampini 98508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 98608122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9879162d606SStefano Zampini cum += subset_size; 9889162d606SStefano Zampini cum2 += subset_size*n; 98908122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 99008122e43SStefano Zampini } 99108122e43SStefano Zampini if (mss) { 9929ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 99308122e43SStefano Zampini PetscBLASInt B_itype = 1; 99408122e43SStefano Zampini PetscBLASInt B_N = mss; 9954c6709b3SStefano Zampini PetscReal zero = 0.0; 9964c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 99708122e43SStefano Zampini 99808122e43SStefano Zampini B_lwork = -1; 99908122e43SStefano Zampini S = NULL; 100008122e43SStefano Zampini St = NULL; 1001a58a30b4SStefano Zampini eigs = NULL; 1002a58a30b4SStefano Zampini eigv = NULL; 1003a58a30b4SStefano Zampini B_iwork = NULL; 1004a58a30b4SStefano Zampini B_ifail = NULL; 1005d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1006d1710679SStefano Zampini rwork = NULL; 1007d1710679SStefano Zampini #endif 10088bec7fa6SStefano Zampini thresh = 1.0; 100908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 101008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 101108122e43SStefano 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)); 101208122e43SStefano Zampini #else 101308122e43SStefano 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)); 101408122e43SStefano Zampini #endif 101508122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 101608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 101708122e43SStefano Zampini } else { 101808122e43SStefano Zampini /* TODO */ 101908122e43SStefano Zampini } 102008122e43SStefano Zampini } else { 102108122e43SStefano Zampini lwork = 0; 102208122e43SStefano Zampini } 102308122e43SStefano Zampini 102408122e43SStefano Zampini nv = 0; 1025d62866d3SStefano 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) */ 1026d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 102708122e43SStefano Zampini } 10284c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1029f6f667cfSStefano Zampini if (allocated_S_St) { 1030f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1031f6f667cfSStefano Zampini } 1032f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 103308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 103408122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 103508122e43SStefano Zampini #endif 10369162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10379162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10389162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 103908122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10409162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 104108122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 104208122e43SStefano Zampini 104308122e43SStefano Zampini maxneigs = 0; 104472b8c272SStefano Zampini cum = cumarray = 0; 10459162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10469162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1047d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 104808122e43SStefano Zampini const PetscInt *idxs; 104908122e43SStefano Zampini 1050d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 105108122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 105208122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 105308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 105408122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10559162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10569162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 105708122e43SStefano Zampini } 1058d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 105908122e43SStefano Zampini } 106008122e43SStefano Zampini 106108122e43SStefano Zampini if (mss) { /* multilevel */ 106208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 106308122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 106408122e43SStefano Zampini } 106508122e43SStefano Zampini 1066ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 106708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 106808122e43SStefano Zampini const PetscInt *idxs; 10699d54b7f4SStefano Zampini PetscReal upper,lower; 1070862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 107108122e43SStefano Zampini PetscBLASInt B_N; 1072aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 107308122e43SStefano Zampini 10749d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 10759d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 10769d54b7f4SStefano Zampini lower = thresh; 10779d54b7f4SStefano Zampini } else { 10789d54b7f4SStefano Zampini upper = 1./thresh; 10799d54b7f4SStefano Zampini lower = 0.; 10809d54b7f4SStefano Zampini } 1081862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1082ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1083f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1084f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10859ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1086aff50787SStefano Zampini PetscInt j,k; 1087aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1088aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1089aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109008122e43SStefano Zampini } 109108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1092aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1093aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1094aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1095aff50787SStefano Zampini } 109608122e43SStefano Zampini } 109708122e43SStefano Zampini } else { 109808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110008122e43SStefano Zampini } 11018bec7fa6SStefano Zampini } else { 1102f6f667cfSStefano Zampini S = Sarray + cumarray; 1103f6f667cfSStefano Zampini St = Starray + cumarray; 11048bec7fa6SStefano Zampini } 1105aff50787SStefano Zampini /* see if we can save some work */ 1106b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1107aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1108aff50787SStefano Zampini } 1109aff50787SStefano Zampini 1110b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1111aff50787SStefano Zampini B_neigs = 0; 1112aff50787SStefano Zampini } else { 11139ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 111408122e43SStefano Zampini PetscBLASInt B_itype = 1; 1115f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11164c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11179552c7c7SStefano Zampini PetscInt nmin_s; 1118b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 111908122e43SStefano Zampini 1120fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11218bec7fa6SStefano 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]]); 1122fd14bc51SStefano Zampini } 1123d16cbb6bSStefano Zampini 1124b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1125b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1126b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1127b7ab4a40SStefano Zampini } 1128b7ab4a40SStefano Zampini 112908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1130b7ab4a40SStefano Zampini if (compute_range) { 1131d16cbb6bSStefano Zampini 1132d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 113308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11349d54b7f4SStefano 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)); 113508122e43SStefano Zampini #else 11369d54b7f4SStefano 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)); 113708122e43SStefano Zampini #endif 1138b7ab4a40SStefano Zampini } else if (!same_data) { 1139d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1140d16cbb6bSStefano Zampini B_IL = 1; 1141d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 11429d54b7f4SStefano 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)); 1143d16cbb6bSStefano Zampini #else 11449d54b7f4SStefano 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)); 1145d16cbb6bSStefano Zampini #endif 1146b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1147b7ab4a40SStefano Zampini PetscInt k; 1148b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1149b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1150b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1151b7ab4a40SStefano Zampini nmin = nmax; 1152b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1153b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1154b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1155b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1156b7ab4a40SStefano Zampini } 1157d16cbb6bSStefano Zampini } 115808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 115908122e43SStefano Zampini if (B_ierr) { 11606c4ed002SBarry 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); 11616c4ed002SBarry 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); 11626c4ed002SBarry 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); 116308122e43SStefano Zampini } 116408122e43SStefano Zampini 116508122e43SStefano Zampini if (B_neigs > nmax) { 1166fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1167fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1168fd14bc51SStefano Zampini } 11699d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 117008122e43SStefano Zampini B_neigs = nmax; 117108122e43SStefano Zampini } 117208122e43SStefano Zampini 11739552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11749552c7c7SStefano Zampini if (B_neigs < nmin_s) { 117508122e43SStefano Zampini PetscBLASInt B_neigs2; 117608122e43SStefano Zampini 11779d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1178f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 11799d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 11809d54b7f4SStefano Zampini } else { 11819d54b7f4SStefano Zampini B_IL = B_neigs + 1; 11829d54b7f4SStefano Zampini B_IU = nmin_s; 11839d54b7f4SStefano Zampini } 1184fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1185fd14bc51SStefano 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); 1186fd14bc51SStefano Zampini } 11879ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 118808122e43SStefano Zampini PetscInt j; 118908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 119008122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 119108122e43SStefano Zampini } 119208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 119308122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 119408122e43SStefano Zampini } 119508122e43SStefano Zampini } else { 119608122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119708122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119808122e43SStefano Zampini } 119908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 120008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 12019d54b7f4SStefano 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)); 120208122e43SStefano Zampini #else 12039d54b7f4SStefano 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)); 120408122e43SStefano Zampini #endif 120508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 120608122e43SStefano Zampini B_neigs += B_neigs2; 120708122e43SStefano Zampini } 120808122e43SStefano Zampini if (B_ierr) { 12096c4ed002SBarry 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); 12106c4ed002SBarry 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); 12116c4ed002SBarry 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); 121208122e43SStefano Zampini } 1213fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1214ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 121508122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 121608122e43SStefano Zampini if (eigs[j] == 0.0) { 1217ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 121808122e43SStefano Zampini } else { 12199d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1220ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 12219d54b7f4SStefano Zampini } else { 12229d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 12239d54b7f4SStefano Zampini } 1224fd14bc51SStefano Zampini } 122508122e43SStefano Zampini } 122608122e43SStefano Zampini } 122708122e43SStefano Zampini } else { 122808122e43SStefano Zampini /* TODO */ 122908122e43SStefano Zampini } 1230aff50787SStefano Zampini } 12316c3e6151SStefano Zampini /* change the basis back to the original one */ 12326c3e6151SStefano Zampini if (sub_schurs->change) { 123372b8c272SStefano Zampini Mat change,phi,phit; 12346c3e6151SStefano Zampini 12356c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12366c3e6151SStefano Zampini PetscInt ii; 12376c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12386c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12396c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12406c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12416c3e6151SStefano Zampini } 12426c3e6151SStefano Zampini } 12436c3e6151SStefano Zampini } 124472b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12456c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 124672b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12476c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12486c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12496c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12506c3e6151SStefano Zampini } 12518bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12528bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12539162d606SStefano Zampini if (B_neigs) { 12549162d606SStefano 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); 1255fd14bc51SStefano Zampini 1256fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12579552c7c7SStefano Zampini PetscInt ii; 12589552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1259ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12609552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1261ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1262ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1263ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1264ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1265ac47001eSStefano Zampini #else 1266ac47001eSStefano 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); 1267ac47001eSStefano Zampini #endif 12689552c7c7SStefano Zampini } 12699552c7c7SStefano Zampini } 1270fd14bc51SStefano Zampini } 12719162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12729162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12739162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12749162d606SStefano Zampini cum++; 127508122e43SStefano Zampini } 127608122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 127708122e43SStefano Zampini /* shift for next computation */ 127808122e43SStefano Zampini cumarray += subset_size*subset_size; 127908122e43SStefano Zampini } 1280fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1281fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1282fd14bc51SStefano Zampini } 128308122e43SStefano Zampini 128408122e43SStefano Zampini if (mss) { 128508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 128608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1287f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1288f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1289f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 129008122e43SStefano Zampini } 1291f6f667cfSStefano Zampini if (allocated_S_St) { 1292f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1293f6f667cfSStefano Zampini } 1294f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 129508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 129608122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 129708122e43SStefano Zampini #endif 129808122e43SStefano Zampini if (pcbddc->dbg_flag) { 12991b968477SStefano Zampini PetscInt maxneigs_r; 1300b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 13019b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 130208122e43SStefano Zampini } 130308122e43SStefano Zampini PetscFunctionReturn(0); 130408122e43SStefano Zampini } 1305b1b3d7a2SStefano Zampini 1306674ae819SStefano Zampini #undef __FUNCT__ 1307c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1308c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1309c8587f34SStefano Zampini { 1310c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13118629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1312c8587f34SStefano Zampini PetscErrorCode ierr; 1313c8587f34SStefano Zampini 1314c8587f34SStefano Zampini PetscFunctionBegin; 1315f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 13165e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1317c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1318c8587f34SStefano Zampini 1319684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 13200fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1321684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1322c8587f34SStefano Zampini 1323c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1324b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1325c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1326c8587f34SStefano Zampini } 1327c8587f34SStefano Zampini 13288629588bSStefano Zampini /* 13298629588bSStefano Zampini Setup local correction and local part of coarse basis. 13308629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13318629588bSStefano Zampini */ 133247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13338629588bSStefano Zampini 13348629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13358629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13368629588bSStefano Zampini 13378629588bSStefano Zampini /* free */ 13388629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1339c8587f34SStefano Zampini PetscFunctionReturn(0); 1340c8587f34SStefano Zampini } 1341c8587f34SStefano Zampini 1342c8587f34SStefano Zampini #undef __FUNCT__ 1343674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1344674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1345674ae819SStefano Zampini { 1346674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1347674ae819SStefano Zampini PetscErrorCode ierr; 1348674ae819SStefano Zampini 1349674ae819SStefano Zampini PetscFunctionBegin; 1350674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1351674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 135230368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1353674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1354674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1355785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1356674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1357f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1358f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1359785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 136063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 136163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1362674ae819SStefano Zampini PetscFunctionReturn(0); 1363674ae819SStefano Zampini } 1364674ae819SStefano Zampini 1365674ae819SStefano Zampini #undef __FUNCT__ 1366674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1367674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1368674ae819SStefano Zampini { 1369674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13704f1b2e48SStefano Zampini PetscInt i; 1371674ae819SStefano Zampini PetscErrorCode ierr; 1372674ae819SStefano Zampini 1373674ae819SStefano Zampini PetscFunctionBegin; 1374b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1375674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 137616909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 13771dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1378674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1379674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13804f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13814f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13824f1b2e48SStefano Zampini } 13834f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1384b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1385674ae819SStefano Zampini PetscFunctionReturn(0); 1386674ae819SStefano Zampini } 1387674ae819SStefano Zampini 1388674ae819SStefano Zampini #undef __FUNCT__ 1389674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1390674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1391674ae819SStefano Zampini { 1392674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1393674ae819SStefano Zampini PetscErrorCode ierr; 1394674ae819SStefano Zampini 1395674ae819SStefano Zampini PetscFunctionBegin; 1396674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 139758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1398ca92afb2SStefano Zampini PetscScalar *array; 139906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 140006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 140158da7f69SStefano Zampini } 1402674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1403674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 140415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 140515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1406674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1407674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1408674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 140906656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1410674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1411674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14128ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1413674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1414674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1415674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1416f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1417f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1418f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1419f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1420727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 14210e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1422f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 142370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 142481d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14250369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14261dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 14274f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14288b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1429ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1430ca92afb2SStefano Zampini PetscInt i; 1431ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1432ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1433ca92afb2SStefano Zampini } 1434ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1435ca92afb2SStefano Zampini } 14364f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1437674ae819SStefano Zampini PetscFunctionReturn(0); 1438674ae819SStefano Zampini } 1439674ae819SStefano Zampini 1440674ae819SStefano Zampini #undef __FUNCT__ 1441f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1442f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14436bfb1811SStefano Zampini { 14446bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14456bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14466bfb1811SStefano Zampini VecType impVecType; 14474f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14486bfb1811SStefano Zampini PetscErrorCode ierr; 14496bfb1811SStefano Zampini 14506bfb1811SStefano Zampini PetscFunctionBegin; 14516c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1452e7b262bdSStefano Zampini /* get sizes */ 14534f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1454b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14556bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1456e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1457e7b262bdSStefano Zampini /* R nodes */ 1458e7b262bdSStefano Zampini old_size = -1; 1459e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1460e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1461e7b262bdSStefano Zampini } 1462e7b262bdSStefano Zampini if (n_R != old_size) { 1463e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1464e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14656bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14666bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14676bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14686bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1469e7b262bdSStefano Zampini } 1470e7b262bdSStefano Zampini /* local primal dofs */ 1471e7b262bdSStefano Zampini old_size = -1; 1472e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1473e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1474e7b262bdSStefano Zampini } 1475e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1476e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 147783b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1478e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14796bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1480e7b262bdSStefano Zampini } 1481e7b262bdSStefano Zampini /* local explicit constraints */ 1482e7b262bdSStefano Zampini old_size = -1; 1483e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1484e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1485e7b262bdSStefano Zampini } 1486e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1487e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 148883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 148983b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 149083b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 149183b7ccabSStefano Zampini } 14926bfb1811SStefano Zampini PetscFunctionReturn(0); 14936bfb1811SStefano Zampini } 14946bfb1811SStefano Zampini 14956bfb1811SStefano Zampini #undef __FUNCT__ 149647f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 149747f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 149888ebb749SStefano Zampini { 149925084f0cSStefano Zampini PetscErrorCode ierr; 150025084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 150188ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 150288ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1503d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 150425084f0cSStefano Zampini /* submatrices of local problem */ 150580677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 150606656605SStefano Zampini /* submatrices of local coarse problem */ 150706656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 150825084f0cSStefano Zampini /* working matrices */ 150906656605SStefano Zampini Mat C_CR; 151025084f0cSStefano Zampini /* additional working stuff */ 151106656605SStefano Zampini PC pc_R; 15124f1b2e48SStefano Zampini Mat F; 15135cbda25cSStefano Zampini Vec dummy_vec; 1514a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 151525084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 151606656605SStefano Zampini PetscScalar *work; 151706656605SStefano Zampini PetscInt *idx_V_B; 1518ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 151906656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1520ffd830a3SStefano Zampini 152125084f0cSStefano Zampini /* some shortcuts to scalars */ 152206656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 152388ebb749SStefano Zampini 152488ebb749SStefano Zampini PetscFunctionBegin; 1525ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1526ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1527ffd830a3SStefano Zampini } 1528ffd830a3SStefano Zampini 1529ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1530b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15314f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1532b371cd4fSStefano Zampini n_B = pcis->n_B; 1533b371cd4fSStefano Zampini n_D = pcis->n - n_B; 153488ebb749SStefano Zampini n_R = pcis->n - n_vertices; 153588ebb749SStefano Zampini 153688ebb749SStefano Zampini /* vertices in boundary numbering */ 1537785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15380e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 15396c4ed002SBarry 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); 154088ebb749SStefano Zampini 154106656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1542019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 154306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 154406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 154506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 154606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 154706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 154806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 154906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 155106656605SStefano Zampini 155206656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 155306656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 155406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 155506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 155606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1557ffd830a3SStefano Zampini lda_rhs = n_R; 1558a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 155906656605SStefano Zampini if (isLU || isILU || isCHOL) { 156006656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1561df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1562df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1563d62866d3SStefano Zampini MatFactorType type; 1564d62866d3SStefano Zampini 1565df4d28bfSStefano Zampini F = reuse_solver->F; 15666816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1567d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1568ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1569df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 157006656605SStefano Zampini } else { 157106656605SStefano Zampini F = NULL; 157206656605SStefano Zampini } 157306656605SStefano Zampini 1574ffd830a3SStefano Zampini /* allocate workspace */ 1575ffd830a3SStefano Zampini n = 0; 1576ffd830a3SStefano Zampini if (n_constraints) { 1577ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1578ffd830a3SStefano Zampini } 1579ffd830a3SStefano Zampini if (n_vertices) { 1580ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1581ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1582ffd830a3SStefano Zampini } 1583ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1584ffd830a3SStefano Zampini 15855cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 15865cbda25cSStefano Zampini dummy_vec = NULL; 15875cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 15885cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 15895cbda25cSStefano Zampini } 15905cbda25cSStefano Zampini 159188ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 159288ebb749SStefano Zampini if (n_constraints) { 159372b8c272SStefano Zampini Mat M1,M2,M3,C_B; 159406656605SStefano Zampini IS is_aux; 159580677318SStefano Zampini PetscScalar *array,*array2; 159606656605SStefano Zampini 1597f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 159880677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 159988ebb749SStefano Zampini 160025084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 160125084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 16028ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 160372b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 160488ebb749SStefano Zampini 160580677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 160680677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1607ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 160888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 160906656605SStefano Zampini const PetscScalar *row_cmat_values; 161006656605SStefano Zampini const PetscInt *row_cmat_indices; 161106656605SStefano Zampini PetscInt size_of_constraint,j; 161288ebb749SStefano Zampini 161306656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161406656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1615ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 161606656605SStefano Zampini } 161706656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161806656605SStefano Zampini } 1619ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 162006656605SStefano Zampini if (F) { 162106656605SStefano Zampini Mat B; 162206656605SStefano Zampini 1623ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1624a3df083aSStefano Zampini if (need_benign_correction) { 1625df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1626a3df083aSStefano Zampini 162772b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 162872b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1629a3df083aSStefano Zampini } 163080677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1631a3df083aSStefano Zampini if (need_benign_correction) { 1632a3df083aSStefano Zampini PetscScalar *marr; 1633df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1634a3df083aSStefano Zampini 1635a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 16365cbda25cSStefano Zampini if (lda_rhs != n_R) { 16375cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 16385cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 16395cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 16405cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 16415cbda25cSStefano Zampini } 16425cbda25cSStefano Zampini } else { 1643a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1644a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 16455cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1646a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1647a3df083aSStefano Zampini } 16485cbda25cSStefano Zampini } 1649a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1650a3df083aSStefano Zampini } 165106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 165206656605SStefano Zampini } else { 165380677318SStefano Zampini PetscScalar *marr; 165480677318SStefano Zampini 165580677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 165606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1657ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1658ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 165906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 166006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 166106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 166206656605SStefano Zampini } 166380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 166406656605SStefano Zampini } 166580677318SStefano Zampini if (!pcbddc->switch_static) { 166680677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 166780677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 166880677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 166980677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1670ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 167180677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 167280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167480677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 167580677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 167680677318SStefano Zampini } 167780677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 167880677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 167972b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 168080677318SStefano Zampini } else { 1681ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1682ffd830a3SStefano Zampini IS dummy; 1683ffd830a3SStefano Zampini 1684ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 168572b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1686ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1687ffd830a3SStefano Zampini } else { 168880677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 168980677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1690ffd830a3SStefano Zampini } 169125084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 169280677318SStefano Zampini } 169380677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 169480677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 169580677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 169606656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 169706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 169880677318SStefano Zampini if (isCHOL) { 169980677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 170080677318SStefano Zampini } else { 170125084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 170280677318SStefano Zampini } 170380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 170406656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 170525084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 170625084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 170725084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 170880677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 170972b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 171072b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 171106656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 171206656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1713f4ddd8eeSStefano Zampini } 1714fc227af8SStefano Zampini 1715fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 171688ebb749SStefano Zampini if (n_vertices) { 171706656605SStefano Zampini IS is_aux; 17183a50541eSStefano Zampini 1719df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 17206816873aSStefano Zampini IS tis; 17216816873aSStefano Zampini 17226816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17236816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17246816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17256816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17266816873aSStefano Zampini } else { 17273a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17286816873aSStefano Zampini } 17299577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17309577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 173104708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 173225084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 173388ebb749SStefano Zampini } 173488ebb749SStefano Zampini 173588ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1736f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 173706656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 173806656605SStefano Zampini if (pcbddc->coarse_phi_D) { 173906656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 174006656605SStefano Zampini } 1741f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 174206656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 174306656605SStefano Zampini PetscScalar *marray; 174406656605SStefano Zampini 174506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 174606656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1747f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1748f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1749f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1750f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1751f4ddd8eeSStefano Zampini } 1752f4ddd8eeSStefano Zampini } 175306656605SStefano Zampini 1754f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 175506656605SStefano Zampini PetscScalar *marray; 175688ebb749SStefano Zampini 175706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17588eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 175906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 176088ebb749SStefano Zampini } 17613301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 176206656605SStefano Zampini n *= 2; 176388ebb749SStefano Zampini } 176406656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 176506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 176606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17678eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 176806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 176906656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 177088ebb749SStefano Zampini } 17713301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 177206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17738eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 177406656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 177506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 177688ebb749SStefano Zampini } 177788ebb749SStefano Zampini } else { 1778c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1779c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17801b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1781c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1782c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1783c0553b1fSStefano Zampini } 178488ebb749SStefano Zampini } 178506656605SStefano Zampini } 1786019a44ceSStefano Zampini 178706656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17884f1b2e48SStefano Zampini p0_lidx_I = NULL; 17894f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1790d12edf2fSStefano Zampini const PetscInt *idxs; 1791d12edf2fSStefano Zampini 1792d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17934f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17944f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17954f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17964f1b2e48SStefano Zampini } 1797d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1798d12edf2fSStefano Zampini } 1799d16cbb6bSStefano Zampini 180006656605SStefano Zampini /* vertices */ 180106656605SStefano Zampini if (n_vertices) { 180216f15bc4SStefano Zampini 1803af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 180404708bb6SStefano Zampini 180516f15bc4SStefano Zampini if (n_R) { 180614393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 180706656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 180816f15bc4SStefano Zampini PetscScalar *x,*y; 180904708bb6SStefano Zampini PetscBool isseqaij; 181006656605SStefano Zampini 181121eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 181214393ed6SStefano Zampini if (need_benign_correction) { 181314393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 181414393ed6SStefano Zampini IS is_p0; 181514393ed6SStefano Zampini PetscInt *idxs_p0,n; 181614393ed6SStefano Zampini 181714393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 181814393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 181914393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 1820af25d912SStefano 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); 182114393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 182214393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 182314393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 182414393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 182514393ed6SStefano Zampini } 182614393ed6SStefano Zampini 1827ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1828af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1829ffd830a3SStefano Zampini } else { 1830ca92afb2SStefano Zampini PetscScalar *av,*array; 1831ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1832ca92afb2SStefano Zampini PetscInt n; 1833ca92afb2SStefano Zampini PetscBool flg_row; 1834ffd830a3SStefano Zampini 1835ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1836ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 18379d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1838ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1839ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1840ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1841ca92afb2SStefano Zampini PetscInt j; 1842ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1843ffd830a3SStefano Zampini } 1844ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1845ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1846ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1847ffd830a3SStefano Zampini } 1848ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1849a3df083aSStefano Zampini if (need_benign_correction) { 1850df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1851a3df083aSStefano Zampini PetscScalar *marr; 1852a3df083aSStefano Zampini 1853a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 185414393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 185514393ed6SStefano Zampini 185614393ed6SStefano Zampini | 0 0 0 | (V) 185714393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 185814393ed6SStefano Zampini | 0 0 -1 | (p0) 185914393ed6SStefano Zampini 186014393ed6SStefano Zampini */ 1861df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 186214393ed6SStefano Zampini const PetscScalar *vals; 186314393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 186414393ed6SStefano Zampini PetscInt n,j,nz; 186514393ed6SStefano Zampini 1866df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1867df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186814393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 186914393ed6SStefano Zampini for (j=0;j<n;j++) { 187014393ed6SStefano Zampini PetscScalar val = vals[j]; 187114393ed6SStefano Zampini PetscInt k,col = idxs[j]; 187214393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 187314393ed6SStefano Zampini } 187414393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1875df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 187614393ed6SStefano Zampini } 187772b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 187872b8c272SStefano Zampini } 187972b8c272SStefano Zampini if (F) { 188014393ed6SStefano Zampini /* need to correct the rhs */ 188172b8c272SStefano Zampini if (need_benign_correction) { 188272b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 188372b8c272SStefano Zampini PetscScalar *marr; 188472b8c272SStefano Zampini 188572b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 18865cbda25cSStefano Zampini if (lda_rhs != n_R) { 18875cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18885cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18895cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 18905cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18915cbda25cSStefano Zampini } 18925cbda25cSStefano Zampini } else { 1893a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1894a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18955cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1896a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1897a3df083aSStefano Zampini } 18985cbda25cSStefano Zampini } 1899a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1900a3df083aSStefano Zampini } 190106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 190214393ed6SStefano Zampini /* need to correct the solution */ 1903a3df083aSStefano Zampini if (need_benign_correction) { 1904df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1905a3df083aSStefano Zampini PetscScalar *marr; 1906a3df083aSStefano Zampini 1907a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 19085cbda25cSStefano Zampini if (lda_rhs != n_R) { 19095cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 19105cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19115cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19125cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19135cbda25cSStefano Zampini } 19145cbda25cSStefano Zampini } else { 1915a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1916a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19175cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1918a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1919a3df083aSStefano Zampini } 19205cbda25cSStefano Zampini } 1921a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1922a3df083aSStefano Zampini } 192306656605SStefano Zampini } else { 192406656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 192506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1926ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1927ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 192806656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 192906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 193006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 193106656605SStefano Zampini } 193206656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 193306656605SStefano Zampini } 193480677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1935ffd830a3SStefano Zampini /* S_VV and S_CV */ 193606656605SStefano Zampini if (n_constraints) { 193706656605SStefano Zampini Mat B; 193880677318SStefano Zampini 1939ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 194080677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1941ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1942ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 194380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 194680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 194780677318SStefano Zampini } 1948ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 194980677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 195080677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1951ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 195280677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 195306656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1954ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1955ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 195606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195706656605SStefano Zampini } 195804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 195904708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 1960511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 196104708bb6SStefano Zampini } 1962ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1963ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1964ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1965ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1966ffd830a3SStefano Zampini } 196706656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 196814393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 196914393ed6SStefano Zampini if (need_benign_correction) { 1970df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 197114393ed6SStefano Zampini PetscScalar *marr,*sums; 197214393ed6SStefano Zampini 197314393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 197414393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1975df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 197614393ed6SStefano Zampini const PetscScalar *vals; 197714393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 197814393ed6SStefano Zampini PetscInt n,j,nz; 197914393ed6SStefano Zampini 1980df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1981df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 198214393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 198314393ed6SStefano Zampini PetscInt k; 198414393ed6SStefano Zampini sums[j] = 0.; 198514393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 198614393ed6SStefano Zampini } 198714393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 198814393ed6SStefano Zampini for (j=0;j<n;j++) { 198914393ed6SStefano Zampini PetscScalar val = vals[j]; 199014393ed6SStefano Zampini PetscInt k; 199114393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 199214393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 199314393ed6SStefano Zampini } 199414393ed6SStefano Zampini } 199514393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1996df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 199714393ed6SStefano Zampini } 199814393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 199914393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 200014393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 200114393ed6SStefano Zampini } 200280677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 200306656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 200406656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 200506656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 200606656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 200706656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 200806656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 200906656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2010d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2011019a44ceSStefano Zampini } else { 2012d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2013d16cbb6bSStefano Zampini } 201421eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2015d16cbb6bSStefano Zampini 201606656605SStefano Zampini /* coarse basis functions */ 201706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 201816f15bc4SStefano Zampini PetscScalar *y; 201916f15bc4SStefano Zampini 2020ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 202106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 202306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 202606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 202806656605SStefano Zampini 202906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20304f1b2e48SStefano Zampini PetscInt j; 20314f1b2e48SStefano Zampini 203206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 203406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20374f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 203806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203906656605SStefano Zampini } 204006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 204106656605SStefano Zampini } 204204708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 204304708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 204406656605SStefano Zampini } 20455cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 204606656605SStefano Zampini 204706656605SStefano Zampini if (n_constraints) { 204806656605SStefano Zampini Mat B; 204906656605SStefano Zampini 2050ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 205106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205280677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 205306656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205406656605SStefano Zampini if (n_vertices) { 205580677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 205680677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 205780677318SStefano Zampini } else { 205880677318SStefano Zampini Mat S_VCt; 205980677318SStefano Zampini 2060ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2061ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206272b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2063ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2064ffd830a3SStefano Zampini } 206580677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 206680677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 206780677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 206880677318SStefano Zampini } 206906656605SStefano Zampini } 207006656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 207106656605SStefano Zampini /* coarse basis functions */ 207206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 207306656605SStefano Zampini PetscScalar *y; 207406656605SStefano Zampini 2075ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 207606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 207806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 208106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 208206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20834f1b2e48SStefano Zampini PetscInt j; 20844f1b2e48SStefano Zampini 208506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 208606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 208706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20904f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 209106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 209206656605SStefano Zampini } 209306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209406656605SStefano Zampini } 209506656605SStefano Zampini } 209680677318SStefano Zampini if (n_constraints) { 209780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 209880677318SStefano Zampini } 20994f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 210072b8c272SStefano Zampini 210172b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 210272b8c272SStefano Zampini if (pcbddc->benign_n) { 210372b8c272SStefano Zampini Mat B0_B,B0_BPHI; 210472b8c272SStefano Zampini IS is_dummy; 210572b8c272SStefano Zampini PetscScalar *data; 210672b8c272SStefano Zampini PetscInt j; 210772b8c272SStefano Zampini 210872b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 210972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 211072b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 211172b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 211286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 211372b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 211472b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 211572b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 211672b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 211772b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 211872b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 211972b8c272SStefano Zampini } 212072b8c272SStefano Zampini } 212172b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 212272b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 212372b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 212472b8c272SStefano Zampini } 2125019a44ceSStefano Zampini 212606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 21273301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2128ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2129ffd830a3SStefano Zampini PetscScalar *marray; 213006656605SStefano Zampini 213106656605SStefano Zampini if (n_constraints) { 2132ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 213306656605SStefano Zampini 2134af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 213506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2136ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 213716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 213806656605SStefano Zampini if (n_vertices) { 2139ffd830a3SStefano Zampini Mat S_VCT; 214006656605SStefano Zampini 214106656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2142ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 214316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 214406656605SStefano Zampini } 2145ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 214606656605SStefano Zampini } 214716f15bc4SStefano Zampini if (n_vertices && n_R) { 2148ffd830a3SStefano Zampini PetscScalar *av,*marray; 2149ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2150ffd830a3SStefano Zampini PetscInt n; 2151ffd830a3SStefano Zampini PetscBool flg_row; 215206656605SStefano Zampini 2153ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2154af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2155ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2156ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2157ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2158ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2159ffd830a3SStefano Zampini PetscInt j; 2160ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2161ffd830a3SStefano Zampini } 2162ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2163ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2164ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 216506656605SStefano Zampini } 216606656605SStefano Zampini 2167ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2168ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2169ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2170ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2171ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 217206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 217306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 217506656605SStefano Zampini } 2176ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2177ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2178ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2179ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2180ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2181ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2182ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2183ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 218406656605SStefano Zampini } 2185ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 218606656605SStefano Zampini /* coarse basis functions */ 218706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 218806656605SStefano Zampini PetscScalar *y; 218906656605SStefano Zampini 2190ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 219106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 219306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219506656605SStefano Zampini if (i<n_vertices) { 219606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 219706656605SStefano Zampini } 219806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 220006656605SStefano Zampini 220106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 220206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 220406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 220706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220806656605SStefano Zampini } 220906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 221006656605SStefano Zampini } 2211ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2212ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 221306656605SStefano Zampini } 2214d62866d3SStefano Zampini /* free memory */ 221588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 221606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 221706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 221806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 221906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2220d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2221d62866d3SStefano Zampini if (n_vertices) { 2222d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2223d62866d3SStefano Zampini } 2224d62866d3SStefano Zampini if (n_constraints) { 2225d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2226d62866d3SStefano Zampini } 222788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 222888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 222988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2230d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 223188ebb749SStefano Zampini Mat coarse_sub_mat; 223225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 223388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 223488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 223588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22368bec7fa6SStefano Zampini Mat C_B,CPHI; 22378bec7fa6SStefano Zampini IS is_dummy; 22388bec7fa6SStefano Zampini Vec mones; 223988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 224088ebb749SStefano Zampini PetscReal real_value; 224188ebb749SStefano Zampini 2242a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2243a3df083aSStefano Zampini Mat A; 2244a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2245a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2246a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2247a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2248a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2249a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2250a3df083aSStefano Zampini } else { 225188ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 225288ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 225388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 225488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2255a3df083aSStefano Zampini } 225688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 225788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2258ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 225988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 226088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 226188ebb749SStefano Zampini } 226288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 226388ebb749SStefano Zampini 226425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22653301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 226625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2267ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 226988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 227088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 227688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 227988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228088ebb749SStefano Zampini } else { 228188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 228288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 228388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 228588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 228888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228988ebb749SStefano Zampini } 229088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2293511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 22944f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2295fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2296d12edf2fSStefano Zampini PetscScalar *data,*data2; 22974f1b2e48SStefano Zampini PetscInt j; 2298d12edf2fSStefano Zampini 22994f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2300fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2301d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 230286c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2303d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2304d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 23054f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 23064f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2307d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 23084f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 23094f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 23104f1b2e48SStefano Zampini } 2311d12edf2fSStefano Zampini } 2312d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2313d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2314d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2315d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2316d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2317d12edf2fSStefano Zampini } 2318d12edf2fSStefano Zampini #if 0 2319d12edf2fSStefano Zampini { 2320d12edf2fSStefano Zampini PetscViewer viewer; 2321d12edf2fSStefano Zampini char filename[256]; 2322ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2323d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2324d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2325ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2326ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2327ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2328d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 232972b8c272SStefano Zampini if (save_change) { 233072b8c272SStefano Zampini Mat phi_B; 233172b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 233272b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 233372b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 233472b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 233572b8c272SStefano Zampini } else { 2336ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2337ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 233872b8c272SStefano Zampini } 2339ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2340ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2341ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2342ffd830a3SStefano Zampini } 2343ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2344ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2345ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2346ffd830a3SStefano Zampini } 234772b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2348ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2349ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2350ffd830a3SStefano Zampini } 2351d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2352d12edf2fSStefano Zampini } 2353d12edf2fSStefano Zampini #endif 235481d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23558bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23561575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 235706656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23588bec7fa6SStefano Zampini 23598bec7fa6SStefano Zampini /* check constraints */ 2360a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2361a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 23624f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23638bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2364a00504b5SStefano Zampini } else { 2365a00504b5SStefano Zampini PetscScalar *data; 2366a00504b5SStefano Zampini Mat tmat; 2367a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2368a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2369a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2370a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2371a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2372a00504b5SStefano Zampini } 23738bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23748bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23758bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23768bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2377bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2378ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2379bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2380bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2381bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2382bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2383bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 238488ebb749SStefano Zampini } 23858bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23868bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23878bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23888bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 238925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 239088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 239188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 239288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 239388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 239488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 239588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 239688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 239788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 239888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 239988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2400ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 240188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 240288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 240388ebb749SStefano Zampini } 240488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 240588ebb749SStefano Zampini } 24068629588bSStefano Zampini /* get back data */ 24078629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 240888ebb749SStefano Zampini PetscFunctionReturn(0); 240988ebb749SStefano Zampini } 241088ebb749SStefano Zampini 241188ebb749SStefano Zampini #undef __FUNCT__ 2412d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2413d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2414aa0d41d4SStefano Zampini { 2415d65f70fdSStefano Zampini Mat *work_mat; 2416d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2417d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2418d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2419aa0d41d4SStefano Zampini PetscErrorCode ierr; 2420aa0d41d4SStefano Zampini 2421aa0d41d4SStefano Zampini PetscFunctionBegin; 2422d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2423d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2424d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2425d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2426aa0d41d4SStefano Zampini 2427d65f70fdSStefano Zampini if (!rsorted) { 2428906d46d4SStefano Zampini const PetscInt *idxs; 2429906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2430aa0d41d4SStefano Zampini 2431d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2432d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2433d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2434d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2435aa0d41d4SStefano Zampini } 2436d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2437d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2438d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2439d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2440aa0d41d4SStefano Zampini } 2441d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2442d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2443d65f70fdSStefano Zampini } else { 2444d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2445d65f70fdSStefano Zampini isrow_s = isrow; 2446aa0d41d4SStefano Zampini } 2447906d46d4SStefano Zampini 2448d65f70fdSStefano Zampini if (!csorted) { 2449d65f70fdSStefano Zampini if (isrow == iscol) { 2450d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2451d65f70fdSStefano Zampini iscol_s = isrow_s; 2452d65f70fdSStefano Zampini } else { 2453d65f70fdSStefano Zampini const PetscInt *idxs; 2454d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2455906d46d4SStefano Zampini 2456d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2457d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2458d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2459d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2460d65f70fdSStefano Zampini } 2461d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2462d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2463d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2464d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2465d65f70fdSStefano Zampini } 2466d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2467d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2468d65f70fdSStefano Zampini } 2469d65f70fdSStefano Zampini } else { 2470d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2471d65f70fdSStefano Zampini iscol_s = iscol; 2472d65f70fdSStefano Zampini } 2473d65f70fdSStefano Zampini 2474d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2475d65f70fdSStefano Zampini 2476d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2477906d46d4SStefano Zampini Mat new_mat; 2478d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2479906d46d4SStefano Zampini 2480d65f70fdSStefano Zampini if (!rsorted) { 2481d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2482d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2483d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2484d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2485906d46d4SStefano Zampini } 2486d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2487d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2488d65f70fdSStefano Zampini } else { 2489d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2490906d46d4SStefano Zampini } 2491d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2492d65f70fdSStefano Zampini 2493d65f70fdSStefano Zampini if (!csorted) { 2494d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2495d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2496d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2497d65f70fdSStefano Zampini } else { 2498d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2499d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2500d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2501d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2502d65f70fdSStefano Zampini } 2503d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2504d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2505d65f70fdSStefano Zampini } 2506d65f70fdSStefano Zampini } else { 2507d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2508d65f70fdSStefano Zampini } 2509d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2510d65f70fdSStefano Zampini 2511d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2512d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2513d65f70fdSStefano Zampini work_mat[0] = new_mat; 2514d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2515d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2516d65f70fdSStefano Zampini } 2517d65f70fdSStefano Zampini 2518d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2519d65f70fdSStefano Zampini *B = work_mat[0]; 2520d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2521d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2522d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2523d65f70fdSStefano Zampini PetscFunctionReturn(0); 2524d65f70fdSStefano Zampini } 2525d65f70fdSStefano Zampini 2526d65f70fdSStefano Zampini #undef __FUNCT__ 25275e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 25285e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2529aa0d41d4SStefano Zampini { 2530aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 25315e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2532d65f70fdSStefano Zampini Mat new_mat; 25335e8657edSStefano Zampini IS is_local,is_global; 2534d65f70fdSStefano Zampini PetscInt local_size; 2535d65f70fdSStefano Zampini PetscBool isseqaij; 2536aa0d41d4SStefano Zampini PetscErrorCode ierr; 2537aa0d41d4SStefano Zampini 2538aa0d41d4SStefano Zampini PetscFunctionBegin; 2539aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25405e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25415e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2542b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2543aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2544d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2545aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2546906d46d4SStefano Zampini 2547906d46d4SStefano Zampini /* check */ 2548906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2549906d46d4SStefano Zampini Vec x,x_change; 2550906d46d4SStefano Zampini PetscReal error; 2551906d46d4SStefano Zampini 25525e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2553906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25545e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2555e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2556e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2557d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2558e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2559e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2560906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2561906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2562906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2563906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2564906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2565906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2566906d46d4SStefano Zampini } 2567906d46d4SStefano Zampini 256822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25699b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 257022d5777bSStefano Zampini if (isseqaij) { 2571a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2572a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2573aa0d41d4SStefano Zampini } else { 2574a00504b5SStefano Zampini Mat work_mat; 25751cf9b237SStefano Zampini 2576a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2577aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2578a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2579aa0d41d4SStefano Zampini } 25803301b35fSStefano Zampini if (matis->A->symmetric_set) { 25813301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2582e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25833301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2584e496cd5dSStefano Zampini #endif 25853301b35fSStefano Zampini } 2586d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2587aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2588aa0d41d4SStefano Zampini } 2589aa0d41d4SStefano Zampini 2590aa0d41d4SStefano Zampini #undef __FUNCT__ 2591a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25928ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2593a64d13efSStefano Zampini { 2594a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2595a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2596d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 259753892102SStefano Zampini PetscInt *idx_R_local=NULL; 25983a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25993a50541eSStefano Zampini PetscInt vbs,bs; 26006816873aSStefano Zampini PetscBT bitmask=NULL; 2601a64d13efSStefano Zampini PetscErrorCode ierr; 2602a64d13efSStefano Zampini 2603a64d13efSStefano Zampini PetscFunctionBegin; 2604b23d619eSStefano Zampini /* 2605b23d619eSStefano Zampini No need to setup local scatters if 2606b23d619eSStefano Zampini - primal space is unchanged 2607b23d619eSStefano Zampini AND 2608b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2609b23d619eSStefano Zampini AND 2610b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2611b23d619eSStefano Zampini */ 2612b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2613f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2614f4ddd8eeSStefano Zampini } 2615f4ddd8eeSStefano Zampini /* destroy old objects */ 2616f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2617f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2618f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2619a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2620b371cd4fSStefano Zampini n_B = pcis->n_B; 2621b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2622b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 26233a50541eSStefano Zampini 2624a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 26256816873aSStefano Zampini 262653892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2627df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2628854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2629a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2630a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 26310e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2632a64d13efSStefano Zampini } 2633a64d13efSStefano Zampini 2634a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26354641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26366816873aSStefano Zampini idx_R_local[n_R++] = i; 2637a64d13efSStefano Zampini } 2638a64d13efSStefano Zampini } 2639df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2640df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26416816873aSStefano Zampini 2642df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2643df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26446816873aSStefano Zampini } 26453a50541eSStefano Zampini 26463a50541eSStefano Zampini /* Block code */ 26473a50541eSStefano Zampini vbs = 1; 26483a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26493a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26503a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26513a50541eSStefano Zampini PetscInt *vary; 2652df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2653785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26543a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2655d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2656d3df7717SStefano 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 */ 26570e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2658d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26593a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26603a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26613a50541eSStefano Zampini break; 26623a50541eSStefano Zampini } 26633a50541eSStefano Zampini } 2664d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2665d3df7717SStefano Zampini } else { 2666d3df7717SStefano Zampini /* Verify directly the R set */ 2667d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2668d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2669d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2670d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2671d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2672d3df7717SStefano Zampini break; 2673d3df7717SStefano Zampini } 2674d3df7717SStefano Zampini } 2675d3df7717SStefano Zampini } 2676d3df7717SStefano Zampini } 26773a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26783a50541eSStefano Zampini vbs = bs; 26793a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26803a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26813a50541eSStefano Zampini } 26823a50541eSStefano Zampini } 26833a50541eSStefano Zampini } 26843a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2685df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2686df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 268753892102SStefano Zampini 2688df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2689df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 269053892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2691df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 269253892102SStefano Zampini } else { 26933a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 269453892102SStefano Zampini } 2695a64d13efSStefano Zampini 2696a64d13efSStefano Zampini /* print some info if requested */ 2697a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2698a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2699a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 27001575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2701a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2702a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 27034f1b2e48SStefano 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); 2704a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2705a64d13efSStefano Zampini } 2706a64d13efSStefano Zampini 2707a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2708df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 27096816873aSStefano Zampini IS is_aux1,is_aux2; 27106816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 27116816873aSStefano Zampini 27123a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2713854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2714854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2715a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 27164641a718SStefano Zampini for (i=0; i<n_D; i++) { 27174641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 27184641a718SStefano Zampini } 2719a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2720a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27214641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 27224641a718SStefano Zampini aux_array1[j++] = i; 2723a64d13efSStefano Zampini } 2724a64d13efSStefano Zampini } 2725a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2726a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2727a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 27284641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 27294641a718SStefano Zampini aux_array2[j++] = i; 2730a64d13efSStefano Zampini } 2731a64d13efSStefano Zampini } 2732a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2733a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2734a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2735a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2736a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2737a64d13efSStefano Zampini 27388eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2739785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2740a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27414641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27424641a718SStefano Zampini aux_array1[j++] = i; 2743a64d13efSStefano Zampini } 2744a64d13efSStefano Zampini } 2745a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2746a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2747a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2748a64d13efSStefano Zampini } 27494641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27503a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2751d62866d3SStefano Zampini } else { 2752df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27536816873aSStefano Zampini IS tis; 27546816873aSStefano Zampini PetscInt schur_size; 27556816873aSStefano Zampini 2756df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27576816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2758df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27596816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27606816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27616816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27626816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27636816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2764d62866d3SStefano Zampini } 2765d62866d3SStefano Zampini } 2766a64d13efSStefano Zampini PetscFunctionReturn(0); 2767a64d13efSStefano Zampini } 2768a64d13efSStefano Zampini 2769304d26faSStefano Zampini 2770304d26faSStefano Zampini #undef __FUNCT__ 2771304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2772684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2773304d26faSStefano Zampini { 2774304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2775304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2776304d26faSStefano Zampini PC pc_temp; 2777304d26faSStefano Zampini Mat A_RR; 2778f4ddd8eeSStefano Zampini MatReuse reuse; 2779304d26faSStefano Zampini PetscScalar m_one = -1.0; 2780304d26faSStefano Zampini PetscReal value; 278104708bb6SStefano Zampini PetscInt n_D,n_R; 27829577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2783304d26faSStefano Zampini PetscErrorCode ierr; 2784e604994aSStefano Zampini /* prefixes stuff */ 2785312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2786e604994aSStefano Zampini size_t len; 2787304d26faSStefano Zampini 2788304d26faSStefano Zampini PetscFunctionBegin; 2789304d26faSStefano Zampini 2790e604994aSStefano Zampini /* compute prefixes */ 2791e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2792e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2793e604994aSStefano Zampini if (!pcbddc->current_level) { 2794e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2795e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2796e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2797e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2798e604994aSStefano Zampini } else { 2799e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2800312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2801e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2802e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2803312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2804312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 280534d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 280634d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2807e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2808e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2809e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2810e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2811e604994aSStefano Zampini } 2812e604994aSStefano Zampini 2813304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2814684f6988SStefano Zampini if (dirichlet) { 2815d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2816450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2817450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2818450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2819450f8f5eSStefano Zampini } 2820450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2821a3df083aSStefano Zampini Mat A_IIn; 2822a3df083aSStefano Zampini 2823a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2824a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2825a3df083aSStefano Zampini pcis->A_II = A_IIn; 2826a3df083aSStefano Zampini } 2827450f8f5eSStefano Zampini } 28283301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28293301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2830964fefecSStefano Zampini } 2831ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2832964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2833304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2834304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2835304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2836304d26faSStefano Zampini /* default */ 2837304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2838e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28399577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2840304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28419577ea80SStefano Zampini if (issbaij) { 28429577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28439577ea80SStefano Zampini } else { 2844304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28459577ea80SStefano Zampini } 2846304d26faSStefano Zampini /* Allow user's customization */ 2847304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2848304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2849304d26faSStefano Zampini } 2850d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2851df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2852df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2853d62866d3SStefano Zampini 2854df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2855d5574798SStefano Zampini } 2856304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2857304d26faSStefano Zampini if (!n_D) { 2858304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2859304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2860304d26faSStefano Zampini } 2861304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2862304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2863304d26faSStefano Zampini /* set ksp_D into pcis data */ 2864304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2865304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2866304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2867684f6988SStefano Zampini } 2868304d26faSStefano Zampini 2869304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2870684f6988SStefano Zampini A_RR = 0; 2871684f6988SStefano Zampini if (neumann) { 2872d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 287304708bb6SStefano Zampini PetscInt ibs,mbs; 287404708bb6SStefano Zampini PetscBool issbaij; 287504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2876f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28778ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2878f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2879f4ddd8eeSStefano Zampini PetscInt nn_R; 288081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2881f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2882f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2883f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2884f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2885f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2886f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2887f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2888727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2889f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2890f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2891f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2892f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2893f4ddd8eeSStefano Zampini } 2894f4ddd8eeSStefano Zampini } 2895f4ddd8eeSStefano Zampini /* last check */ 2896d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2897f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2898f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2899f4ddd8eeSStefano Zampini } 2900f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2901f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2902f4ddd8eeSStefano Zampini } 2903a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 2904af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2905af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 290604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 290704708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 290804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 290904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2911af732b37SStefano Zampini } else { 2912511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 29136816873aSStefano Zampini } 291404708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 291504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 291804708bb6SStefano Zampini } else { 2919511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 292004708bb6SStefano Zampini } 292104708bb6SStefano Zampini } 2922a00504b5SStefano Zampini /* extract A_RR */ 2923a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 2924a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2925a00504b5SStefano Zampini 2926a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 292716e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2928a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 292916e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 293016e386b8SStefano Zampini } else { 2931a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 2932a00504b5SStefano Zampini } 2933a00504b5SStefano Zampini } else { 2934a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2935a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 2936a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2937a00504b5SStefano Zampini } 2938a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 2939f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 294016e386b8SStefano Zampini } 29413301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29423301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29436816873aSStefano Zampini } 2944f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2945304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2946304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2947304d26faSStefano Zampini /* default */ 2948304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2949e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2950304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29519577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29529577ea80SStefano Zampini if (issbaij) { 29539577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29549577ea80SStefano Zampini } else { 2955304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29569577ea80SStefano Zampini } 2957304d26faSStefano Zampini /* Allow user's customization */ 2958304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2959304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2960304d26faSStefano Zampini } 2961304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2962304d26faSStefano Zampini if (!n_R) { 2963304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2964304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2965304d26faSStefano Zampini } 29665cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2967df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2968df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2969df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2970d62866d3SStefano Zampini 2971df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2972d62866d3SStefano Zampini } 2973304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2974304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2975684f6988SStefano Zampini } 2976304d26faSStefano Zampini 2977304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29780fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2979684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2980684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29811575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2982684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2983684f6988SStefano Zampini } 2984684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29850fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29860fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29870fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29880fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29890fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2990304d26faSStefano Zampini /* need to be adapted? */ 2991b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2992b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2993b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2994304d26faSStefano Zampini /* print info */ 2995304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2996e604994aSStefano 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); 2997304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2998304d26faSStefano Zampini } 2999b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 3000298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 3001304d26faSStefano Zampini } 3002684f6988SStefano Zampini } 3003684f6988SStefano Zampini if (neumann) { /* Neumann */ 30040fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 30050fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 30060fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 30070fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 30080fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3009304d26faSStefano Zampini /* need to be adapted? */ 3010b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 3011b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3012304d26faSStefano Zampini /* print info */ 3013304d26faSStefano Zampini if (pcbddc->dbg_flag) { 3014e604994aSStefano 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); 3015304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3016304d26faSStefano Zampini } 3017b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 3018298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 3019304d26faSStefano Zampini } 30200fccc4e9SStefano Zampini } 3021684f6988SStefano Zampini } 30225cbda25cSStefano Zampini /* free Neumann problem's matrix */ 30235cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3024304d26faSStefano Zampini PetscFunctionReturn(0); 3025304d26faSStefano Zampini } 3026304d26faSStefano Zampini 3027304d26faSStefano Zampini #undef __FUNCT__ 3028ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 302980677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3030674ae819SStefano Zampini { 3031674ae819SStefano Zampini PetscErrorCode ierr; 3032674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3033be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3034674ae819SStefano Zampini 3035674ae819SStefano Zampini PetscFunctionBegin; 3036df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 303780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 303820c7b377SStefano Zampini } 303980677318SStefano Zampini if (!pcbddc->switch_static) { 304080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 304180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 304280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 304320c7b377SStefano Zampini } 3044df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 304580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304720c7b377SStefano Zampini } else { 3048df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3049be83ff47SStefano Zampini 3050df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3051df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305220c7b377SStefano Zampini } 3053be83ff47SStefano Zampini } else { 305480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 305980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 306080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 306180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3063674ae819SStefano Zampini } 3064674ae819SStefano Zampini } 306572b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 306680677318SStefano Zampini if (applytranspose) { 306780677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 306880677318SStefano Zampini } else { 306980677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 307080677318SStefano Zampini } 3071be83ff47SStefano Zampini } else { 3072df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3073be83ff47SStefano Zampini 3074be83ff47SStefano Zampini if (applytranspose) { 3075df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3076be83ff47SStefano Zampini } else { 3077df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3078be83ff47SStefano Zampini } 3079be83ff47SStefano Zampini } 308080677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 308180677318SStefano Zampini if (!pcbddc->switch_static) { 3082df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 308380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3085be83ff47SStefano Zampini } else { 3086df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3087be83ff47SStefano Zampini 3088df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3089df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3090be83ff47SStefano Zampini } 309180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 309280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 309380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 309480677318SStefano Zampini } 309580677318SStefano Zampini } else { 309680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 310180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 310280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 310380677318SStefano Zampini } 310480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3108674ae819SStefano Zampini } 3109674ae819SStefano Zampini PetscFunctionReturn(0); 3110674ae819SStefano Zampini } 3111674ae819SStefano Zampini 3112dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3113674ae819SStefano Zampini #undef __FUNCT__ 3114674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3115dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3116674ae819SStefano Zampini { 3117674ae819SStefano Zampini PetscErrorCode ierr; 3118674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3119674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3120674ae819SStefano Zampini const PetscScalar zero = 0.0; 3121674ae819SStefano Zampini 3122674ae819SStefano Zampini PetscFunctionBegin; 3123dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 31244fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3125dc359a40SStefano Zampini if (applytranspose) { 3126674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 31278eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3128dc359a40SStefano Zampini } else { 3129674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3130674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 313115aaf578SStefano Zampini } 31324fee134fSStefano Zampini } else { 31334fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 31344fee134fSStefano Zampini } 3135efc2fbd9SStefano Zampini 3136efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31374f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3138efc2fbd9SStefano Zampini PetscScalar *array; 31394f1b2e48SStefano Zampini PetscInt j; 3140efc2fbd9SStefano Zampini 3141efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31424f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3143efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3144efc2fbd9SStefano Zampini } 3145efc2fbd9SStefano Zampini 314612edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 314712edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 314812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 314912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315012edc857SStefano Zampini 31519f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 315212edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 315312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 315451694757SStefano Zampini Mat coarse_mat; 3155964fefecSStefano Zampini Vec rhs,sol; 315651694757SStefano Zampini MatNullSpace nullsp; 3157964fefecSStefano Zampini 3158964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3159964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 316051694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 316151694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 316251694757SStefano Zampini if (nullsp) { 316351694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 316451694757SStefano Zampini } 316512edc857SStefano Zampini if (applytranspose) { 3166964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 316712edc857SStefano Zampini } else { 3168964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 316912edc857SStefano Zampini } 317051694757SStefano Zampini if (nullsp) { 317151694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 317251694757SStefano Zampini } 317312edc857SStefano Zampini } 3174674ae819SStefano Zampini 3175674ae819SStefano Zampini /* Local solution on R nodes */ 31764fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 317780677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31789f00e9b4SStefano Zampini } 31799f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31809f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 318112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3182674ae819SStefano Zampini 31834fee134fSStefano Zampini /* Sum contributions from the two levels */ 31844fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3185dc359a40SStefano Zampini if (applytranspose) { 3186dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3187dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3188dc359a40SStefano Zampini } else { 3189674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31908eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3191dc359a40SStefano Zampini } 3192efc2fbd9SStefano Zampini /* store p0 */ 31934f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3194efc2fbd9SStefano Zampini PetscScalar *array; 31954f1b2e48SStefano Zampini PetscInt j; 3196efc2fbd9SStefano Zampini 3197efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31984f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3199efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3200efc2fbd9SStefano Zampini } 32014fee134fSStefano Zampini } else { /* expand the coarse solution */ 32024fee134fSStefano Zampini if (applytranspose) { 32034fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32044fee134fSStefano Zampini } else { 32054fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32064fee134fSStefano Zampini } 32074fee134fSStefano Zampini } 3208674ae819SStefano Zampini PetscFunctionReturn(0); 3209674ae819SStefano Zampini } 3210674ae819SStefano Zampini 3211674ae819SStefano Zampini #undef __FUNCT__ 3212674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 321312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3214674ae819SStefano Zampini { 3215674ae819SStefano Zampini PetscErrorCode ierr; 3216674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 321758da7f69SStefano Zampini PetscScalar *array; 321812edc857SStefano Zampini Vec from,to; 3219674ae819SStefano Zampini 3220674ae819SStefano Zampini PetscFunctionBegin; 322112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 322212edc857SStefano Zampini from = pcbddc->coarse_vec; 322312edc857SStefano Zampini to = pcbddc->vec1_P; 322412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 322512edc857SStefano Zampini Vec tvec; 322658da7f69SStefano Zampini 322758da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 322858da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 322912edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 323058da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 323158da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 323258da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 323312edc857SStefano Zampini } 323412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 323512edc857SStefano Zampini from = pcbddc->vec1_P; 323612edc857SStefano Zampini to = pcbddc->coarse_vec; 323712edc857SStefano Zampini } 323812edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3239674ae819SStefano Zampini PetscFunctionReturn(0); 3240674ae819SStefano Zampini } 3241674ae819SStefano Zampini 3242674ae819SStefano Zampini #undef __FUNCT__ 3243674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 324412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3245674ae819SStefano Zampini { 3246674ae819SStefano Zampini PetscErrorCode ierr; 3247674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 324858da7f69SStefano Zampini PetscScalar *array; 324912edc857SStefano Zampini Vec from,to; 3250674ae819SStefano Zampini 3251674ae819SStefano Zampini PetscFunctionBegin; 325212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 325312edc857SStefano Zampini from = pcbddc->coarse_vec; 325412edc857SStefano Zampini to = pcbddc->vec1_P; 325512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 325612edc857SStefano Zampini from = pcbddc->vec1_P; 325712edc857SStefano Zampini to = pcbddc->coarse_vec; 325812edc857SStefano Zampini } 325912edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 326012edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 326112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 326212edc857SStefano Zampini Vec tvec; 326358da7f69SStefano Zampini 326412edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 326558da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 326658da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 326758da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 326858da7f69SStefano Zampini } 326958da7f69SStefano Zampini } else { 327058da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 327158da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 327212edc857SStefano Zampini } 327312edc857SStefano Zampini } 3274674ae819SStefano Zampini PetscFunctionReturn(0); 3275674ae819SStefano Zampini } 3276674ae819SStefano Zampini 3277984c4197SStefano Zampini /* uncomment for testing purposes */ 3278984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3279674ae819SStefano Zampini #undef __FUNCT__ 3280674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3281674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3282674ae819SStefano Zampini { 3283674ae819SStefano Zampini PetscErrorCode ierr; 3284674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3285674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3286674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3287984c4197SStefano Zampini /* one and zero */ 3288984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3289984c4197SStefano Zampini /* space to store constraints and their local indices */ 32909162d606SStefano Zampini PetscScalar *constraints_data; 32919162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32929162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32939162d606SStefano Zampini PetscInt *constraints_n; 3294984c4197SStefano Zampini /* iterators */ 3295b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3296984c4197SStefano Zampini /* BLAS integers */ 3297e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3298e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3299c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3300727cdba6SStefano Zampini /* reuse */ 33010e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 33020e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3303984c4197SStefano Zampini /* change of basis */ 3304b3d85658SStefano Zampini PetscBool qr_needed; 33059162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3306984c4197SStefano Zampini /* auxiliary stuff */ 330764efe560SStefano Zampini PetscInt *nnz,*is_indices; 33088a0068c3SStefano Zampini PetscInt ncc; 3309984c4197SStefano Zampini /* some quantities */ 331045a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3311a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3312984c4197SStefano Zampini 3313674ae819SStefano Zampini PetscFunctionBegin; 33148e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 33158e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 33168e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 331716909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3318088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3319088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 33200e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 33210e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 33220e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33230e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33240e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3325088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3326cf5a6209SStefano Zampini 3327cf5a6209SStefano Zampini /* print some info */ 3328cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3329cf5a6209SStefano Zampini IS vertices; 3330cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3331cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3332cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3333cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 33341575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3335cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3336cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3337fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3338fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3339cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33401575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3341cf5a6209SStefano Zampini } 3342cf5a6209SStefano Zampini 3343cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33449162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3345cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3346cf5a6209SStefano Zampini const Vec *nearnullvecs; 3347cf5a6209SStefano Zampini Vec *localnearnullsp; 3348cf5a6209SStefano Zampini PetscScalar *array; 3349cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3350cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3351674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3352b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3353674ae819SStefano Zampini PetscScalar *work; 3354674ae819SStefano Zampini PetscReal *singular_vals; 3355674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3356674ae819SStefano Zampini PetscReal *rwork; 3357674ae819SStefano Zampini #endif 3358674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3359674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3360674ae819SStefano Zampini #else 3361964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3362964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3363674ae819SStefano Zampini #endif 3364674ae819SStefano Zampini 3365674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3366d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3367d06fc5fdSStefano Zampini /* free unneeded index sets */ 3368d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3369d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3370674ae819SStefano Zampini } 3371d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3372d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3373d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3374d06fc5fdSStefano Zampini } 3375d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3376d06fc5fdSStefano Zampini n_ISForEdges = 0; 3377d06fc5fdSStefano Zampini } 3378d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3379d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3380d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3381d06fc5fdSStefano Zampini } 3382d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3383d06fc5fdSStefano Zampini n_ISForFaces = 0; 3384d06fc5fdSStefano Zampini } 338570022509SStefano Zampini 338670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 338770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 338870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 338970022509SStefano Zampini if (pcbddc->NullSpace) { 339070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 339170022509SStefano Zampini 339270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3393b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3394d06fc5fdSStefano Zampini if (!ISForEdges) { 3395d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3396d06fc5fdSStefano Zampini } 3397b8ffe317SStefano Zampini } 3398d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3399d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3400b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3401d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3402d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 340398a51de6SStefano Zampini } 340470022509SStefano Zampini #endif 340508122e43SStefano Zampini 3406674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3407674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3408674ae819SStefano Zampini if (nearnullsp) { 3409674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3410f4ddd8eeSStefano Zampini /* remove any stored info */ 3411f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3412f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3413f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3414f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3415f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3416473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3417f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3418f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3419f4ddd8eeSStefano Zampini } 3420984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3421984c4197SStefano Zampini nnsp_size = 0; 3422674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3423674ae819SStefano Zampini } 3424984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3425984c4197SStefano Zampini max_constraints = nnsp_size; 3426984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3427984c4197SStefano Zampini 3428674ae819SStefano Zampini /* 3429674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 34309162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 34319162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 34329162d606SStefano Zampini There can be multiple constraints per connected component 3433674ae819SStefano Zampini */ 3434674ae819SStefano Zampini n_vertices = 0; 3435674ae819SStefano Zampini if (ISForVertices) { 3436674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3437674ae819SStefano Zampini } 34389162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34399162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34409162d606SStefano Zampini 34419162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34429162d606SStefano Zampini total_counts *= max_constraints; 3443674ae819SStefano Zampini total_counts += n_vertices; 34444641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34459162d606SStefano Zampini 3446674ae819SStefano Zampini total_counts = 0; 3447674ae819SStefano Zampini max_size_of_constraint = 0; 3448674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34499162d606SStefano Zampini IS used_is; 3450674ae819SStefano Zampini if (i<n_ISForEdges) { 34519162d606SStefano Zampini used_is = ISForEdges[i]; 3452674ae819SStefano Zampini } else { 34539162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3454674ae819SStefano Zampini } 34559162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3456674ae819SStefano Zampini total_counts += j; 3457674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3458674ae819SStefano Zampini } 34599162d606SStefano 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); 34609162d606SStefano Zampini 3461984c4197SStefano Zampini /* get local part of global near null space vectors */ 3462785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3463984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3464984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3465e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3466e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3467984c4197SStefano Zampini } 3468674ae819SStefano Zampini 3469242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3470242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3471a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3472242a89d7SStefano Zampini 3473984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3474a773dcb8SStefano Zampini if (!skip_lapack) { 3475674ae819SStefano Zampini PetscScalar temp_work; 3476911cabfeSStefano Zampini 3477674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3478984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3479785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3480785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3481785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3482674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3483785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3484674ae819SStefano Zampini #endif 3485674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3486c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3487c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3488674ae819SStefano Zampini lwork = -1; 3489674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3490674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3491c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3492674ae819SStefano Zampini #else 3493c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3494674ae819SStefano Zampini #endif 3495674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3496984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3497674ae819SStefano Zampini #else /* on missing GESVD */ 3498674ae819SStefano Zampini /* SVD */ 3499674ae819SStefano Zampini PetscInt max_n,min_n; 3500674ae819SStefano Zampini max_n = max_size_of_constraint; 3501984c4197SStefano Zampini min_n = max_constraints; 3502984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3503674ae819SStefano Zampini min_n = max_size_of_constraint; 3504984c4197SStefano Zampini max_n = max_constraints; 3505674ae819SStefano Zampini } 3506785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3507674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3508785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3509674ae819SStefano Zampini #endif 3510674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3511674ae819SStefano Zampini lwork = -1; 3512e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3513e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3514b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3515674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3516674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35179162d606SStefano 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)); 3518674ae819SStefano Zampini #else 35199162d606SStefano 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)); 3520674ae819SStefano Zampini #endif 3521674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3522984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3523984c4197SStefano Zampini #endif /* on missing GESVD */ 3524674ae819SStefano Zampini /* Allocate optimal workspace */ 3525674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3526854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3527674ae819SStefano Zampini } 3528674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3529674ae819SStefano Zampini total_counts = 0; 35309162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 35319162d606SStefano Zampini constraints_data_ptr[0] = 0; 3532674ae819SStefano Zampini /* vertices */ 35339162d606SStefano Zampini if (n_vertices) { 3534674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 35359162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3536674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35379162d606SStefano Zampini constraints_n[total_counts] = 1; 35389162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35399162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35409162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3541674ae819SStefano Zampini total_counts++; 3542674ae819SStefano Zampini } 3543674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3544674ae819SStefano Zampini n_vertices = total_counts; 3545674ae819SStefano Zampini } 3546984c4197SStefano Zampini 3547674ae819SStefano Zampini /* edges and faces */ 35489162d606SStefano Zampini total_counts_cc = total_counts; 3549911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35509162d606SStefano Zampini IS used_is; 35519162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35529162d606SStefano Zampini 3553911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35549162d606SStefano Zampini used_is = ISForEdges[ncc]; 3555984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3556674ae819SStefano Zampini } else { 35579162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3558984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3559674ae819SStefano Zampini } 3560674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35619162d606SStefano Zampini 35629162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35639162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3564984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3565984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3566674ae819SStefano Zampini if (nnsp_has_cnst) { 35675b08dc53SStefano Zampini PetscScalar quad_value; 35689162d606SStefano Zampini 35699162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35709162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35719162d606SStefano Zampini 3572a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3573674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3574a773dcb8SStefano Zampini } else { 3575a773dcb8SStefano Zampini quad_value = 1.0; 3576a773dcb8SStefano Zampini } 3577674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35789162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3579674ae819SStefano Zampini } 35809162d606SStefano Zampini temp_constraints++; 3581674ae819SStefano Zampini total_counts++; 3582674ae819SStefano Zampini } 3583674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3584984c4197SStefano Zampini PetscReal real_value; 35859162d606SStefano Zampini PetscScalar *ptr_to_data; 35869162d606SStefano Zampini 3587984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35889162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3589674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35909162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3591674ae819SStefano Zampini } 3592984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3593984c4197SStefano Zampini /* check if array is null on the connected component */ 3594e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35959162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35965b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3597674ae819SStefano Zampini temp_constraints++; 3598674ae819SStefano Zampini total_counts++; 35999162d606SStefano Zampini if (!idxs_copied) { 36009162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 36019162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3602674ae819SStefano Zampini } 3603674ae819SStefano Zampini } 36049162d606SStefano Zampini } 36059162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 360645a1bb75SStefano Zampini valid_constraints = temp_constraints; 3607eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3608a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 36099162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 36109162d606SStefano Zampini 36119162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3612a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 36139162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3614a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 36159162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3616a773dcb8SStefano Zampini } else { /* perform SVD */ 3617984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 36189162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3619674ae819SStefano Zampini 3620674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3621984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3622984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3623984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3624984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3625984c4197SStefano Zampini from that computed using LAPACKgesvd 3626984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3627984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3628984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3629674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3630e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3631984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3632674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3633674ae819SStefano Zampini for (k=0;k<j+1;k++) { 36349162d606SStefano 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)); 3635674ae819SStefano Zampini } 3636674ae819SStefano Zampini } 3637e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3638e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3639e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3640674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3641c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3642674ae819SStefano Zampini #else 3643c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3644674ae819SStefano Zampini #endif 3645674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3646984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3647984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3648674ae819SStefano Zampini j = 0; 3649984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3650674ae819SStefano Zampini total_counts = total_counts-j; 365145a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3652e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3653c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3654c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3655c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3656c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3657c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3658c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3659674ae819SStefano Zampini if (j<temp_constraints) { 3660984c4197SStefano Zampini PetscInt ii; 3661984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3662674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36639162d606SStefano 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)); 3664674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3665984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3666674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36679162d606SStefano 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]; 3668674ae819SStefano Zampini } 3669674ae819SStefano Zampini } 3670674ae819SStefano Zampini } 3671674ae819SStefano Zampini #else /* on missing GESVD */ 3672e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3673e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3674b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3675674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3676674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36779162d606SStefano 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)); 3678674ae819SStefano Zampini #else 36799162d606SStefano 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)); 3680674ae819SStefano Zampini #endif 3681984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3682674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3683984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3684e310c8b4SStefano Zampini k = temp_constraints; 3685e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3686674ae819SStefano Zampini j = 0; 3687e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 368845a1bb75SStefano Zampini valid_constraints = k-j; 3689911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3690984c4197SStefano Zampini #endif /* on missing GESVD */ 3691674ae819SStefano Zampini } 3692a773dcb8SStefano Zampini } 36939162d606SStefano Zampini /* update pointers information */ 36949162d606SStefano Zampini if (valid_constraints) { 36959162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36969162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36979162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36989162d606SStefano Zampini /* set change_of_basis flag */ 369945a1bb75SStefano Zampini if (boolforchange) { 3700b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 37019162d606SStefano Zampini } 3702b3d85658SStefano Zampini total_counts_cc++; 370345a1bb75SStefano Zampini } 370445a1bb75SStefano Zampini } 3705984c4197SStefano Zampini /* free workspace */ 37068f1c130eSStefano Zampini if (!skip_lapack) { 3707984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3708984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3709984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3710984c4197SStefano Zampini #endif 3711984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3712984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3713984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3714984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3715984c4197SStefano Zampini #endif 3716984c4197SStefano Zampini } 3717984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3718984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3719984c4197SStefano Zampini } 3720984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3721cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3722cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3723cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3724cf5a6209SStefano Zampini } 3725cf5a6209SStefano Zampini if (n_ISForFaces) { 3726cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3727cf5a6209SStefano Zampini } 3728cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3729cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3730cf5a6209SStefano Zampini } 3731cf5a6209SStefano Zampini if (n_ISForEdges) { 3732cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3733cf5a6209SStefano Zampini } 3734cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 373508122e43SStefano Zampini } else { 373608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3737984c4197SStefano Zampini 373808122e43SStefano Zampini total_counts = 0; 373908122e43SStefano Zampini n_vertices = 0; 3740d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3741d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 374208122e43SStefano Zampini } 374308122e43SStefano Zampini max_constraints = 0; 37449162d606SStefano Zampini total_counts_cc = 0; 374508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 374608122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37479162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 374808122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 374908122e43SStefano Zampini } 37509162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37519162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37529162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37539162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 375474d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37559162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37569162d606SStefano Zampini total_counts_cc = 0; 37579162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37589162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37599162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 376008122e43SStefano Zampini } 376108122e43SStefano Zampini } 37629162d606SStefano Zampini #if 0 37639162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37649162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37659162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37669162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37679162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37689162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37699162d606SStefano Zampini } 37709162d606SStefano Zampini printf("\n"); 37719162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37729162d606SStefano Zampini } 37731b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37748bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37751b968477SStefano Zampini } 37761b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37778bec7fa6SStefano 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]); 37781b968477SStefano Zampini } 377908122e43SStefano Zampini #endif 378008122e43SStefano Zampini 37818bec7fa6SStefano Zampini max_size_of_constraint = 0; 37829162d606SStefano 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]); 37839162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 378408122e43SStefano Zampini /* Change of basis */ 3785b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 378608122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 378708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 378808122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3789b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 379008122e43SStefano Zampini } 379108122e43SStefano Zampini } 379208122e43SStefano Zampini } 379308122e43SStefano Zampini } 3794984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37954f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 379608122e43SStefano Zampini 37979162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37989162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 37996c4ed002SBarry 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); 3800674ae819SStefano Zampini 3801674ae819SStefano Zampini /* Create constraint matrix */ 3802674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 380316f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3804984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3805984c4197SStefano Zampini 3806984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3807a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3808a717540cSStefano Zampini qr_needed = PETSC_FALSE; 380974d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3810984c4197SStefano Zampini total_primal_vertices=0; 3811b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 38129162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 38139162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 381472b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 38159162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3816b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 381764efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 38189162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38199162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3820a717540cSStefano Zampini } 3821b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 382291af6908SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single) { 3823a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3824a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3825a717540cSStefano Zampini } 3826fa434743SStefano Zampini } else { 3827b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3828fa434743SStefano Zampini } 3829a717540cSStefano Zampini } 3830b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3831b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3832674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 383370022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3834b3d85658SStefano Zampini 38354f1b2e48SStefano 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); 38360e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38370e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3838984c4197SStefano Zampini 3839984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 384074d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3841785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3842984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 384374d5cdf7SStefano Zampini 3844984c4197SStefano Zampini j = total_primal_vertices; 384574d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3846b3d85658SStefano Zampini cum = total_primal_vertices; 38479162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38484641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3849b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3850b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3851b3d85658SStefano Zampini cum++; 38529162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 385374d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 385474d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 385574d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 385674d5cdf7SStefano Zampini } 38579162d606SStefano Zampini j += constraints_n[i]; 3858674ae819SStefano Zampini } 3859674ae819SStefano Zampini } 3860674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3861674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3862088faed8SStefano Zampini 3863674ae819SStefano Zampini /* set values in constraint matrix */ 3864984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38650e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3866674ae819SStefano Zampini } 3867984c4197SStefano Zampini total_counts = total_primal_vertices; 38689162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38694641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38709162d606SStefano Zampini PetscInt *cols; 38719162d606SStefano Zampini 38729162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38739162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38749162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38759162d606SStefano Zampini PetscInt row = total_counts+k; 38769162d606SStefano Zampini PetscScalar *vals; 38779162d606SStefano Zampini 38789162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38799162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38809162d606SStefano Zampini } 38819162d606SStefano Zampini total_counts += constraints_n[i]; 3882674ae819SStefano Zampini } 3883674ae819SStefano Zampini } 3884674ae819SStefano Zampini /* assembling */ 3885674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3886674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3887088faed8SStefano Zampini 3888984c4197SStefano Zampini /* 38896a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3890984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3891f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 3892984c4197SStefano Zampini */ 3893674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3894674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3895026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3896984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3897984c4197SStefano Zampini /* working stuff for GEQRF */ 389881d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3899984c4197SStefano Zampini PetscBLASInt lqr_work; 3900984c4197SStefano Zampini /* working stuff for UNGQR */ 3901984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3902984c4197SStefano Zampini PetscBLASInt lgqr_work; 3903984c4197SStefano Zampini /* working stuff for TRTRS */ 3904984c4197SStefano Zampini PetscScalar *trs_rhs; 39053f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3906984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3907984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3908984c4197SStefano Zampini PetscScalar *start_vals; 3909984c4197SStefano Zampini /* working stuff for values insertion */ 39104641a718SStefano Zampini PetscBT is_primal; 391164efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3912906d46d4SStefano Zampini /* matrix sizes */ 3913906d46d4SStefano Zampini PetscInt global_size,local_size; 3914906d46d4SStefano Zampini /* temporary change of basis */ 3915906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3916cf5a6209SStefano Zampini /* extra space for debugging */ 3917cf5a6209SStefano Zampini PetscScalar *dbg_work; 3918984c4197SStefano Zampini 3919906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3920906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 392116f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3922bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3923906d46d4SStefano Zampini /* nonzeros for local mat */ 3924bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 39251dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3926bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 39271dd7afcfSStefano Zampini } else { 39281dd7afcfSStefano Zampini const PetscInt *ii; 39291dd7afcfSStefano Zampini PetscInt n; 39301dd7afcfSStefano Zampini PetscBool flg_row; 39311dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39321dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 39331dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39341dd7afcfSStefano Zampini } 39359162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3936a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39379162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3938a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39399162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3940a717540cSStefano Zampini } else { 39419162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39429162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3943a717540cSStefano Zampini } 3944a717540cSStefano Zampini } 3945a717540cSStefano Zampini } 3946906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3947bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 39481dd7afcfSStefano Zampini /* Set interior change in the matrix */ 39491dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3950bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3951906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3952a717540cSStefano Zampini } 39531dd7afcfSStefano Zampini } else { 39541dd7afcfSStefano Zampini const PetscInt *ii,*jj; 39551dd7afcfSStefano Zampini PetscScalar *aa; 39561dd7afcfSStefano Zampini PetscInt n; 39571dd7afcfSStefano Zampini PetscBool flg_row; 39581dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39591dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39601dd7afcfSStefano Zampini for (i=0;i<n;i++) { 39611dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 39621dd7afcfSStefano Zampini } 39631dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39641dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39651dd7afcfSStefano Zampini } 3966a717540cSStefano Zampini 3967a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3968a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3969a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3970a717540cSStefano Zampini } 3971a717540cSStefano Zampini 3972a717540cSStefano Zampini 3973a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3974a717540cSStefano Zampini /* 3975a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3976a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3977a717540cSStefano Zampini 3978a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3979a717540cSStefano Zampini 3980a6b551f4SStefano 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) 3981a6b551f4SStefano Zampini 3982a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3983a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3984a717540cSStefano Zampini | ... | 3985a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3986a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3987a717540cSStefano Zampini 3988a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3989a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3990a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3991a6b551f4SStefano Zampini 3992a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3993a717540cSStefano Zampini */ 3994a717540cSStefano Zampini if (qr_needed) { 3995984c4197SStefano Zampini /* space to store Q */ 3996854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3997984c4197SStefano Zampini /* first we issue queries for optimal work */ 39983f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39993f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 40003f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4001984c4197SStefano Zampini lqr_work = -1; 40023f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4003984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4004984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4005785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4006984c4197SStefano Zampini lgqr_work = -1; 40073f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 40083f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 40093f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 40103f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40113f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 40123f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4013984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4014984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4015785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4016984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4017785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4018984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4019785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4020a717540cSStefano Zampini /* allocating workspace for check */ 4021a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4022cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4023a717540cSStefano Zampini } 4024a717540cSStefano Zampini } 4025984c4197SStefano Zampini /* array to store whether a node is primal or not */ 40264641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4027473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 40280e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 40296c4ed002SBarry 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); 403039e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 403139e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 403239e2fb2aSStefano Zampini } 403339e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4034984c4197SStefano Zampini 4035a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 40369162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40379162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40384641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4039984c4197SStefano Zampini /* get constraint info */ 40409162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4041984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4042984c4197SStefano Zampini 4043984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40449162d606SStefano 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); 4045674ae819SStefano Zampini } 4046984c4197SStefano Zampini 4047fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4048a717540cSStefano Zampini 4049a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4050a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40519162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4052a717540cSStefano Zampini } 4053984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40549162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4055984c4197SStefano Zampini 4056984c4197SStefano Zampini /* compute QR decomposition of constraints */ 40573f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40583f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40593f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4060674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40613f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4062984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4063674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4064984c4197SStefano Zampini 4065984c4197SStefano Zampini /* explictly compute R^-T */ 4066984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4067984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40683f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40693f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40703f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40713f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4072984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40733f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4074984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4075984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4076984c4197SStefano Zampini 4077a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40783f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40793f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40803f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40813f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4082984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40833f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4084984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4085984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4086984c4197SStefano Zampini 4087984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4088984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4089984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40903f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40913f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40923f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40933f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40943f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40953f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4096984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40979162d606SStefano 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)); 4098984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40999162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4100984c4197SStefano Zampini 4101984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 41029162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4103984c4197SStefano Zampini /* insert cols for primal dofs */ 4104984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4105984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 41069162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4107906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4108984c4197SStefano Zampini } 4109984c4197SStefano Zampini /* insert cols for dual dofs */ 4110984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 41119162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4112984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 41139162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4114906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4115984c4197SStefano Zampini j++; 4116674ae819SStefano Zampini } 4117674ae819SStefano Zampini } 4118984c4197SStefano Zampini 4119984c4197SStefano Zampini /* check change of basis */ 4120984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4121984c4197SStefano Zampini PetscInt ii,jj; 4122984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4123c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4124c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4125c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4126c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4127c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4128c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4129984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4130cf5a6209SStefano 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)); 4131984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4132984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4133984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4134cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4135cf5a6209SStefano 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; 4136674ae819SStefano Zampini } 4137674ae819SStefano Zampini } 4138984c4197SStefano Zampini if (!valid_qr) { 413922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4140984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4141984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4142cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4143cf5a6209SStefano 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])); 4144674ae819SStefano Zampini } 4145cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4146cf5a6209SStefano 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])); 4147984c4197SStefano Zampini } 4148984c4197SStefano Zampini } 4149984c4197SStefano Zampini } 4150674ae819SStefano Zampini } else { 415122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4152674ae819SStefano Zampini } 4153674ae819SStefano Zampini } 4154a717540cSStefano Zampini } else { /* simple transformation block */ 4155a717540cSStefano Zampini PetscInt row,col; 4156a6b551f4SStefano Zampini PetscScalar val,norm; 4157a6b551f4SStefano Zampini 4158a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41599162d606SStefano 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)); 4160a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 41619162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 41629162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4163bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 41649162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4165906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41669162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4167a717540cSStefano Zampini } else { 4168a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41699162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4170a717540cSStefano Zampini if (row != col) { 41719162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4172a717540cSStefano Zampini } else { 41739162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4174a717540cSStefano Zampini } 4175906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4176a717540cSStefano Zampini } 4177a717540cSStefano Zampini } 4178a717540cSStefano Zampini } 417998a51de6SStefano Zampini if (pcbddc->dbg_flag) { 418022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4181a717540cSStefano Zampini } 4182674ae819SStefano Zampini } 4183984c4197SStefano Zampini } else { 4184984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41859162d606SStefano 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); 4186674ae819SStefano Zampini } 4187674ae819SStefano Zampini } 4188674ae819SStefano Zampini } 4189a717540cSStefano Zampini 4190a717540cSStefano Zampini /* free workspace */ 4191a717540cSStefano Zampini if (qr_needed) { 4192984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4193cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4194984c4197SStefano Zampini } 4195984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4196984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4197984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4198984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4199984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4200674ae819SStefano Zampini } 4201a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4202906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4203906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4204906d46d4SStefano Zampini 4205906d46d4SStefano Zampini /* assembling of global change of variable */ 420688c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4207bbb9e6c6SStefano Zampini Mat tmat; 420816f15bc4SStefano Zampini PetscInt bs; 420916f15bc4SStefano Zampini 4210906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4211906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4212bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4213bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4214bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4215bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 421616f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 421716f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4218906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4219bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4220bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4221bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4222bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4223bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4224e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4225e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4226bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4227bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 422888c03ad3SStefano Zampini 4229906d46d4SStefano Zampini /* check */ 4230906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4231906d46d4SStefano Zampini PetscReal error; 4232906d46d4SStefano Zampini Vec x,x_change; 4233906d46d4SStefano Zampini 4234906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4235906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4236906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4237906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4238e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4239e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4240bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4241e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4242e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4243906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4244906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4245906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4246906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4247bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4248906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4249906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4250906d46d4SStefano Zampini } 4251b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4252b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4253b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4254bf3a8328SStefano Zampini 4255bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4256bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4257bf3a8328SStefano Zampini } 4258b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4259ac632422SStefano Zampini Mat S_new,tmat; 4260bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4261bbb9e6c6SStefano Zampini 4262bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 42636816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4264bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4265bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4266bf3a8328SStefano Zampini IS is_V; 4267b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4268b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4269b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4270b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4271b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4272bf3a8328SStefano Zampini } 4273bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4274ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4275b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4276ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4277bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4278bf3a8328SStefano Zampini const PetscScalar *array; 4279bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4280bf3a8328SStefano Zampini PetscInt i,n_V; 4281bf3a8328SStefano Zampini 4282b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4283b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4284b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4285b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4286b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4287b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4288b087196eSStefano Zampini PetscScalar val; 4289b087196eSStefano Zampini PetscInt idx; 4290b087196eSStefano Zampini 4291b087196eSStefano Zampini idx = idxs_V[i]; 4292b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4293b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4294b087196eSStefano Zampini } 4295b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4296b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4297bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4298bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4299bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4300bf3a8328SStefano Zampini } 4301ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4302ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4303ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4304ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4305b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4306ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4307bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4308b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4309bf3a8328SStefano Zampini } 4310ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4311ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4312ac632422SStefano Zampini } 4313b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 431488c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4315b96c3477SStefano Zampini } 4316c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4317c9db6a07SStefano Zampini if (sub_schurs->change) { 4318c9db6a07SStefano Zampini PetscInt i; 4319c9db6a07SStefano Zampini 4320c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4321c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4322c9db6a07SStefano Zampini } 4323c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4324c9db6a07SStefano Zampini } 4325b96c3477SStefano Zampini } 432616909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 432716909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 432816909a7fSStefano Zampini } else { 4329906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 433016909a7fSStefano Zampini } 43311dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 43321dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43331dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43341dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->change_interior,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43351dd7afcfSStefano Zampini } 433672b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 433772b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 433872b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 433972b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 434072b8c272SStefano Zampini } 43411dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 43421dd7afcfSStefano Zampini PetscBool needglobal = PETSC_FALSE; 43431dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43441dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43451dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&needglobal,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43461dd7afcfSStefano Zampini } 43471dd7afcfSStefano Zampini if (!needglobal && pcbddc->user_ChangeOfBasisMatrix) { 4348b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4349b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4350906d46d4SStefano Zampini } else { 43511dd7afcfSStefano Zampini Mat benign_global = NULL; 43521dd7afcfSStefano Zampini if (needglobal) { 43531dd7afcfSStefano Zampini Mat tmat; 43541dd7afcfSStefano Zampini 43551dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 43561dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 43571dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 43581dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43591dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43601dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 43611dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43621dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43631dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 43641dd7afcfSStefano Zampini if (pcbddc->benign_change) { 43651dd7afcfSStefano Zampini Mat M; 43661dd7afcfSStefano Zampini 43671dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 43681dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 43691dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 43701dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4371906d46d4SStefano Zampini } else { 43721dd7afcfSStefano Zampini Mat eye; 43731dd7afcfSStefano Zampini PetscScalar *array; 43741dd7afcfSStefano Zampini 43751dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43761dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 43771dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 43781dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4379906d46d4SStefano Zampini } 43801dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43811dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43821dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43831dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 43841dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 43851dd7afcfSStefano Zampini } 43861dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 43871dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 43881dd7afcfSStefano Zampini } 43891dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 43901dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 43911dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 43921dd7afcfSStefano Zampini } else if (needglobal) { 43931dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 43941dd7afcfSStefano Zampini } 43951dd7afcfSStefano Zampini } 439616909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 439716909a7fSStefano Zampini IS is_global; 439816909a7fSStefano Zampini const PetscInt *gidxs; 439916909a7fSStefano Zampini 440016909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 440116909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 440216909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 440316909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 440416909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 440516909a7fSStefano Zampini } 44061dd7afcfSStefano Zampini } 44071dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 44081dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4409b9b85e73SStefano Zampini } 4410a717540cSStefano Zampini 441172b8c272SStefano Zampini if (!pcbddc->fake_change) { 44124f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 44134f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 44144f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 44154f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4416019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4417019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4418019a44ceSStefano Zampini pcbddc->local_primal_size++; 4419019a44ceSStefano Zampini } 4420019a44ceSStefano Zampini 4421019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4422727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4423727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 44249f47a83aSStefano 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); 4425c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 44260e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 44279f47a83aSStefano 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); 4428727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4429727cdba6SStefano Zampini } 44300e6343abSStefano Zampini } 4431727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4432b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 443372b8c272SStefano Zampini } 443472b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4435727cdba6SStefano Zampini 4436a717540cSStefano Zampini /* flush dbg viewer */ 4437b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4438b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4439b8ffe317SStefano Zampini } 4440a717540cSStefano Zampini 4441e310c8b4SStefano Zampini /* free workspace */ 4442a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 44434641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 444408122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 44459162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 44469162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 444708122e43SStefano Zampini } else { 44489162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 44499162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 44509162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 445108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 445208122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 44539162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 44549162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 445508122e43SStefano Zampini } 4456674ae819SStefano Zampini PetscFunctionReturn(0); 4457674ae819SStefano Zampini } 4458674ae819SStefano Zampini 4459674ae819SStefano Zampini #undef __FUNCT__ 4460674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4461674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4462674ae819SStefano Zampini { 4463674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4464674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4465674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 446614f95afaSStefano Zampini PetscInt ierr,i,N; 4467674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4468674ae819SStefano Zampini 4469674ae819SStefano Zampini PetscFunctionBegin; 44708e61c736SStefano Zampini /* Reset previously computed graph */ 44718e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4472674ae819SStefano Zampini /* Init local Graph struct */ 44737fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 44743bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4475674ae819SStefano Zampini 4476575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 44775099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 44785099eff2SStefano 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); 4479575ad6abSStefano Zampini } 44809577ea80SStefano Zampini 4481674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4482d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 44834d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 44844d379d7bSStefano Zampini PetscInt nvtxs; 4485e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4486674ae819SStefano Zampini 44872fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 44882fffb893SStefano Zampini if (flg_row) { 44894d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4490b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 44912fffb893SStefano Zampini } 44922fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4493674ae819SStefano Zampini } 44949b28b941SStefano Zampini if (pcbddc->dbg_flag) { 44959b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4496674ae819SStefano Zampini } 4497674ae819SStefano Zampini 449863602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 449963602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 450063602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 450195ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 450263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4503e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 450463602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4505674ae819SStefano Zampini } 450663602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 450763602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 450863602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4509674ae819SStefano Zampini } 451063602bcaSStefano Zampini } else { 451163602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 451263602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4513854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 451463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 451563602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 451663602bcaSStefano Zampini } 451763602bcaSStefano Zampini } 4518674ae819SStefano Zampini } 4519674ae819SStefano Zampini 4520674ae819SStefano Zampini /* Setup of Graph */ 4521785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4522e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4523785d1243SStefano Zampini } 4524785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4525e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4526785d1243SStefano Zampini } 452730368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 452830368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 452930368db7SStefano Zampini } 453014f95afaSStefano 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); 4531674ae819SStefano Zampini 45324f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 45334f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 45344f1b2e48SStefano Zampini PetscInt *local_subs; 45354f1b2e48SStefano Zampini 45364f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 45374f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 45384f1b2e48SStefano Zampini const PetscInt *idxs; 45394f1b2e48SStefano Zampini PetscInt nl,j; 45404f1b2e48SStefano Zampini 45414f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 45424f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45434f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 45444f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 45454f1b2e48SStefano Zampini } 45464f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45474f1b2e48SStefano Zampini } 45484f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 45494f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 45504f1b2e48SStefano Zampini } 45514f1b2e48SStefano Zampini 4552674ae819SStefano Zampini /* Graph's connected components analysis */ 4553674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4554674ae819SStefano Zampini 4555674ae819SStefano Zampini /* print some info to stdout */ 4556674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4557302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4558674ae819SStefano Zampini } 4559674ae819SStefano Zampini PetscFunctionReturn(0); 4560674ae819SStefano Zampini } 4561674ae819SStefano Zampini 4562dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4563674ae819SStefano Zampini #undef __FUNCT__ 4564674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4565dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4566674ae819SStefano Zampini { 4567dc456d91SStefano Zampini PetscSF sf; 4568dc456d91SStefano Zampini PetscLayout map; 4569dc456d91SStefano Zampini const PetscInt *idxs; 4570dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4571dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4572dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4573dc456d91SStefano Zampini PetscMPIInt commsize; 4574674ae819SStefano Zampini PetscBool first_found; 4575674ae819SStefano Zampini PetscErrorCode ierr; 4576674ae819SStefano Zampini 4577674ae819SStefano Zampini PetscFunctionBegin; 4578dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4579dc456d91SStefano Zampini if (subset_mult) { 4580dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4581dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4582dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4583674ae819SStefano Zampini } 4584dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4585dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4586dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4587dc456d91SStefano Zampini for (i=0;i<n;i++) { 4588dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4589dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4590674ae819SStefano Zampini } 4591dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4592b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4593dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4594dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4595dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4596dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4597dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4598dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4599dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4600dc456d91SStefano Zampini 4601dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4602dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4603dc456d91SStefano Zampini if (subset_mult) { 4604dc456d91SStefano Zampini const PetscInt* idxs_mult; 4605dc456d91SStefano Zampini 4606dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4607dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4608dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4609674ae819SStefano Zampini } else { 4610dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4611674ae819SStefano Zampini } 4612dc456d91SStefano Zampini /* local size of new subset */ 4613dc456d91SStefano Zampini n_n = 0; 4614dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4615dc456d91SStefano Zampini 4616dc456d91SStefano Zampini /* global indexes in layout */ 4617dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4618dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4619dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4620dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4621dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4622dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4623dc456d91SStefano Zampini 4624dc456d91SStefano Zampini /* reduce from leaves to roots */ 4625dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 462664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 462764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4628dc456d91SStefano Zampini 4629dc456d91SStefano Zampini /* count indexes in local part of layout */ 4630674ae819SStefano Zampini nlocals = 0; 4631674ae819SStefano Zampini first_index = -1; 4632674ae819SStefano Zampini first_found = PETSC_FALSE; 4633dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4634dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4635674ae819SStefano Zampini first_found = PETSC_TRUE; 4636674ae819SStefano Zampini first_index = i; 4637674ae819SStefano Zampini } 4638dc456d91SStefano Zampini nlocals += root_data[i]; 4639674ae819SStefano Zampini } 4640dc456d91SStefano Zampini 4641dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 46425fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4643dc456d91SStefano Zampini start = 0; 464464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46455fa240b1SStefano Zampini #else 464664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46475fa240b1SStefano Zampini start = start-nlocals; 46485fa240b1SStefano Zampini #endif 46495fa240b1SStefano Zampini 4650dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4651dc456d91SStefano Zampini *N_n = start + nlocals; 4652dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4653dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4654674ae819SStefano Zampini } 46555fa240b1SStefano Zampini 46565fa240b1SStefano Zampini /* adapt root data with cumulative */ 4657674ae819SStefano Zampini if (first_found) { 4658dc456d91SStefano Zampini PetscInt old_index; 4659dc456d91SStefano Zampini 4660dc456d91SStefano Zampini root_data[first_index] += start; 4661674ae819SStefano Zampini old_index = first_index; 4662dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4663dc456d91SStefano Zampini if (root_data[i]) { 4664dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4665674ae819SStefano Zampini old_index = i; 4666674ae819SStefano Zampini } 4667674ae819SStefano Zampini } 4668674ae819SStefano Zampini } 4669dc456d91SStefano Zampini 4670dc456d91SStefano Zampini /* from roots to leaves */ 4671dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4672dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4673dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4674dc456d91SStefano Zampini 4675dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4676dc456d91SStefano Zampini if (subset_mult) { 4677dc456d91SStefano Zampini const PetscInt* idxs_mult; 4678dc456d91SStefano Zampini PetscInt cum; 4679dc456d91SStefano Zampini 4680dc456d91SStefano Zampini cum = 0; 4681dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4682dc456d91SStefano Zampini for (i=0;i<n;i++) { 4683dc456d91SStefano Zampini PetscInt j; 4684dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4685674ae819SStefano Zampini } 4686dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4687674ae819SStefano Zampini } else { 4688dc456d91SStefano Zampini for (i=0;i<n;i++) { 4689dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4690674ae819SStefano Zampini } 4691674ae819SStefano Zampini } 4692dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4693dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4694674ae819SStefano Zampini PetscFunctionReturn(0); 4695674ae819SStefano Zampini } 46969a7d3425SStefano Zampini 46979a7d3425SStefano Zampini #undef __FUNCT__ 46989a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 46999a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 47009a7d3425SStefano Zampini { 47019a7d3425SStefano Zampini PetscInt i,j; 47029a7d3425SStefano Zampini PetscScalar *alphas; 47039a7d3425SStefano Zampini PetscErrorCode ierr; 47049a7d3425SStefano Zampini 47059a7d3425SStefano Zampini PetscFunctionBegin; 47069a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4707785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 47089a7d3425SStefano Zampini for (i=0;i<n;i++) { 47099a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 47109a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 47119a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 47129a7d3425SStefano Zampini } 47139a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 47149a7d3425SStefano Zampini PetscFunctionReturn(0); 47159a7d3425SStefano Zampini } 47169a7d3425SStefano Zampini 4717e7931f94SStefano Zampini #undef __FUNCT__ 471870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 471957de7509SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void) 4720e7931f94SStefano Zampini { 472157de7509SStefano Zampini Mat A; 4722e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4723e7931f94SStefano Zampini PetscMPIInt size,rank,color; 472452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 472552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 472657de7509SStefano Zampini PetscInt im_active,active_procs,n,i,local_size,threshold=0; 472757de7509SStefano Zampini PetscInt void_procs,*procs_candidates = NULL; 472857de7509SStefano Zampini PetscBool ismatis,use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4729e7931f94SStefano Zampini PetscSubcomm subcomm; 473052e5ac9dSStefano Zampini PetscErrorCode ierr; 4731a57a6d2fSStefano Zampini 4732e7931f94SStefano Zampini PetscFunctionBegin; 473357de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 473457de7509SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 473557de7509SStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 473657de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,*n_subdomains,2); 473757de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,redprocs,3); 473857de7509SStefano Zampini if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %d\n",*n_subdomains); 473957de7509SStefano Zampini 474057de7509SStefano Zampini if (have_void) *have_void = PETSC_FALSE; 474157de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 474257de7509SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 474357de7509SStefano Zampini ierr = MatISGetLocalMat(mat,&A);CHKERRQ(ierr); 474457de7509SStefano Zampini ierr = MatGetLocalSize(A,&n,NULL);CHKERRQ(ierr); 474557de7509SStefano Zampini im_active = !!(n); 474657de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 474757de7509SStefano Zampini void_procs = size - active_procs; 474857de7509SStefano Zampini /* get ranks of of non-active processes in mat communicator */ 474957de7509SStefano Zampini if (void_procs) { 475057de7509SStefano Zampini PetscInt ncand; 475157de7509SStefano Zampini 475257de7509SStefano Zampini if (have_void) *have_void = PETSC_TRUE; 475357de7509SStefano Zampini ierr = PetscMalloc1(size,&procs_candidates);CHKERRQ(ierr); 475457de7509SStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 475557de7509SStefano Zampini for (i=0,ncand=0;i<size;i++) { 475657de7509SStefano Zampini if (!procs_candidates[i]) { 475757de7509SStefano Zampini procs_candidates[ncand++] = i; 475857de7509SStefano Zampini } 475957de7509SStefano Zampini } 476057de7509SStefano Zampini /* force n_subdomains to be not greater that the number of non-active processes */ 476157de7509SStefano Zampini *n_subdomains = PetscMin(void_procs,*n_subdomains); 476257de7509SStefano Zampini } 476357de7509SStefano Zampini 476457de7509SStefano Zampini /* number of subdomains requested greater than active processes -> just shift the matrix */ 476557de7509SStefano Zampini if (active_procs < *n_subdomains) { 476657de7509SStefano Zampini PetscInt issize,isidx; 476757de7509SStefano Zampini if (im_active) { 476857de7509SStefano Zampini issize = 1; 476957de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 477057de7509SStefano Zampini isidx = procs_candidates[rank]; 477157de7509SStefano Zampini } else { 477257de7509SStefano Zampini isidx = rank; 477357de7509SStefano Zampini } 477457de7509SStefano Zampini } else { 477557de7509SStefano Zampini issize = 0; 477657de7509SStefano Zampini isidx = -1; 477757de7509SStefano Zampini } 477857de7509SStefano Zampini *n_subdomains = active_procs; 477957de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);CHKERRQ(ierr); 478057de7509SStefano Zampini PetscFunctionReturn(0); 478157de7509SStefano Zampini } 4782c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 4783c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4784c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4785e7931f94SStefano Zampini 4786e7931f94SStefano Zampini /* Get info on mapping */ 47873bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 47883bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4789e7931f94SStefano Zampini 4790e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4791785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4792e7931f94SStefano Zampini xadj[0] = 0; 4793e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4794785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4795785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4796e7931f94SStefano Zampini 47972b510759SStefano Zampini if (threshold) { 4798d023bfaeSStefano Zampini PetscInt xadj_count = 0; 47992b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4800d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4801d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4802d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4803d023bfaeSStefano Zampini xadj_count++; 4804e7931f94SStefano Zampini } 4805e7931f94SStefano Zampini } 4806d023bfaeSStefano Zampini xadj[1] = xadj_count; 4807c8587f34SStefano Zampini } else { 4808e7931f94SStefano Zampini if (xadj[1]) { 4809e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4810e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4811c8587f34SStefano Zampini } 4812e7931f94SStefano Zampini } 48133bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4814e7931f94SStefano Zampini if (use_square) { 4815e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4816e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4817e7931f94SStefano Zampini } 4818e7931f94SStefano Zampini } 4819e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4820e7931f94SStefano Zampini 48213837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4822e7931f94SStefano Zampini 4823e7931f94SStefano Zampini /* 4824e7931f94SStefano Zampini Restrict work on active processes only. 4825e7931f94SStefano Zampini */ 4826e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4827e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 48282b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4829d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4830e7931f94SStefano Zampini if (color) { 4831e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4832e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4833e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4834c8587f34SStefano Zampini } else { 483552e5ac9dSStefano Zampini Mat subdomain_adj; 483652e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 483752e5ac9dSStefano Zampini MatPartitioning partitioner; 483852e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 483952e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 484057de7509SStefano Zampini PetscMPIInt size; 4841b0c7d250SStefano Zampini PetscBool aggregate; 4842b0c7d250SStefano Zampini 4843306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4844785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4845e7931f94SStefano Zampini prank = rank; 4846306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 48478002ef2cSStefano Zampini /* 4848e7931f94SStefano Zampini for (i=0;i<size;i++) { 4849e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4850c8587f34SStefano Zampini } 48518002ef2cSStefano Zampini */ 4852e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4853e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4854c8587f34SStefano Zampini } 4855e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4856b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4857b0c7d250SStefano Zampini if (aggregate) { 4858b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4859b0c7d250SStefano Zampini PetscMPIInt nrank; 4860b0c7d250SStefano Zampini PetscScalar *vals; 4861b0c7d250SStefano Zampini 4862b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4863b0c7d250SStefano Zampini lrows = 0; 4864b0c7d250SStefano Zampini if (nrank<redprocs) { 4865b0c7d250SStefano Zampini lrows = size/redprocs; 4866b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4867b0c7d250SStefano Zampini } 48685fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4869b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4870b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4871b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4872b0c7d250SStefano Zampini row = nrank; 4873b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4874b0c7d250SStefano Zampini cols = adjncy; 4875b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4876b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4877b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4878b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4879b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 488052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 488152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 488252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4883b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4884b0c7d250SStefano Zampini } else { 4885306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4886b0c7d250SStefano Zampini } 488722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4888e7931f94SStefano Zampini 4889e7931f94SStefano Zampini /* Partition */ 4890306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4891e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4892e7931f94SStefano Zampini if (use_vwgt) { 48933837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4894e7931f94SStefano Zampini v_wgt[0] = local_size; 4895e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4896c8587f34SStefano Zampini } 489757de7509SStefano Zampini *n_subdomains = PetscMin((PetscInt)size,*n_subdomains); 489857de7509SStefano Zampini ierr = MatPartitioningSetNParts(partitioner,*n_subdomains);CHKERRQ(ierr); 4899e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4900e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 490122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4902e7931f94SStefano Zampini 490352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 490452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 490552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 490652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 490757de7509SStefano Zampini if (!aggregate) { 490857de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 490957de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]]; 491057de7509SStefano Zampini } else { 4911b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 491257de7509SStefano Zampini } 491328143c3dSStefano Zampini } else { 4914b0c7d250SStefano Zampini PetscInt idxs[1]; 4915b0c7d250SStefano Zampini PetscMPIInt tag; 4916b0c7d250SStefano Zampini MPI_Request *reqs; 4917b0c7d250SStefano Zampini 4918b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4919b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4920b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4921b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 492228143c3dSStefano Zampini } 4923b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4924b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4925b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 492657de7509SStefano Zampini if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */ 492757de7509SStefano Zampini ranks_send_to_idx[0] = procs_candidates[oldranks[idxs[0]]]; 492857de7509SStefano Zampini } else { 4929b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4930e7931f94SStefano Zampini } 493157de7509SStefano Zampini } 493252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4933e7931f94SStefano Zampini /* clean up */ 4934e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 493552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4936e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4937e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4938e7931f94SStefano Zampini } 4939e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 494057de7509SStefano Zampini ierr = PetscFree(procs_candidates);CHKERRQ(ierr); 4941e7931f94SStefano Zampini 4942e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4943e7931f94SStefano Zampini i = 1; 4944e7931f94SStefano Zampini if (color) i=0; 494557de7509SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);CHKERRQ(ierr); 4946e7931f94SStefano Zampini PetscFunctionReturn(0); 4947e7931f94SStefano Zampini } 4948e7931f94SStefano Zampini 4949e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4950e7931f94SStefano Zampini 4951e7931f94SStefano Zampini #undef __FUNCT__ 4952e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 495357de7509SStefano 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[]) 4954e7931f94SStefano Zampini { 495570cf5478SStefano Zampini Mat local_mat; 4956e7931f94SStefano Zampini IS is_sends_internal; 49579d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 495828143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 49599d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4960e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4961e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4962e7931f94SStefano Zampini const PetscInt* is_indices; 4963e7931f94SStefano Zampini MatType new_local_type; 4964e7931f94SStefano Zampini /* buffers */ 4965e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 496628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 49679d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4968e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4969e7931f94SStefano Zampini /* MPI */ 497028143c3dSStefano Zampini MPI_Comm comm,comm_n; 497128143c3dSStefano Zampini PetscSubcomm subcomm; 4972e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 497328143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 497428143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 497528143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 497628143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 497728143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4978e7931f94SStefano Zampini PetscErrorCode ierr; 4979e7931f94SStefano Zampini 4980e7931f94SStefano Zampini PetscFunctionBegin; 498157de7509SStefano Zampini PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 4982e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 498328143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 498457de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 498557de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 498657de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_full,5); 498757de7509SStefano Zampini PetscValidLogicalCollectiveBool(mat,reuse,6); 498857de7509SStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,8); 498957de7509SStefano Zampini 499057de7509SStefano Zampini /* further checks */ 4991e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4992e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4993e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4994e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4995e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 499657de7509SStefano Zampini if (reuse && *mat_n) { 499770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 499857de7509SStefano Zampini PetscValidHeaderSpecific(*mat_n,MAT_CLASSID,7); 499970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 500028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 500170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 500270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 500370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 500470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 500570cf5478SStefano Zampini } 5006e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 5007e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 500857de7509SStefano Zampini 5009e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 5010e7931f94SStefano Zampini if (!is_sends) { 501128143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 501257de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);CHKERRQ(ierr); 5013c8587f34SStefano Zampini } else { 5014e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 5015e7931f94SStefano Zampini is_sends_internal = is_sends; 5016c8587f34SStefano Zampini } 5017e7931f94SStefano Zampini 5018e7931f94SStefano Zampini /* get comm */ 5019a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 5020e7931f94SStefano Zampini 5021e7931f94SStefano Zampini /* compute number of sends */ 5022e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 5023e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 5024e7931f94SStefano Zampini 5025e7931f94SStefano Zampini /* compute number of receives */ 5026e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 5027785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 5028e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 5029e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5030e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 5031e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 5032e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 5033e7931f94SStefano Zampini 503428143c3dSStefano Zampini /* restrict comm if requested */ 503528143c3dSStefano Zampini subcomm = 0; 503628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 503728143c3dSStefano Zampini if (restrict_comm) { 5038779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 5039779c1cceSStefano Zampini 504028143c3dSStefano Zampini color = 0; 504153a05cb3SStefano Zampini if (restrict_full) { 504253a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 504353a05cb3SStefano Zampini } else { 504453a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 504553a05cb3SStefano Zampini } 5046b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 504728143c3dSStefano Zampini subcommsize = commsize - subcommsize; 504828143c3dSStefano Zampini /* check if reuse has been requested */ 504957de7509SStefano Zampini if (reuse) { 505028143c3dSStefano Zampini if (*mat_n) { 505128143c3dSStefano Zampini PetscMPIInt subcommsize2; 505228143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 505328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 505428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 505528143c3dSStefano Zampini } else { 505628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 505728143c3dSStefano Zampini } 505828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 5059779c1cceSStefano Zampini PetscMPIInt rank; 5060779c1cceSStefano Zampini 5061779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 506228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 506328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 506428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5065306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 506628143c3dSStefano Zampini } 506728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 506828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 506928143c3dSStefano Zampini } else { 507028143c3dSStefano Zampini comm_n = comm; 507128143c3dSStefano Zampini } 507228143c3dSStefano Zampini 5073e7931f94SStefano Zampini /* prepare send/receive buffers */ 5074785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5075e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5076785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5077e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 507828143c3dSStefano Zampini if (nis) { 5079854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 508028143c3dSStefano Zampini } 5081e7931f94SStefano Zampini 508228143c3dSStefano Zampini /* Get data from local matrices */ 50836c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5084e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5085e7931f94SStefano Zampini /* 5086e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5087e7931f94SStefano Zampini send_buffer_idxs should contain: 5088e7931f94SStefano Zampini - MatType_PRIVATE type 5089e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5090e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5091e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5092e7931f94SStefano Zampini */ 50936c4ed002SBarry Smith else { 5094e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 50953bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5096854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5097e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5098e7931f94SStefano Zampini send_buffer_idxs[1] = i; 50993bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5100e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 51013bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5102e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5103e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5104e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5105e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5106c8587f34SStefano Zampini } 5107c8587f34SStefano Zampini } 5108e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 510928143c3dSStefano Zampini /* additional is (if any) */ 511028143c3dSStefano Zampini if (nis) { 511128143c3dSStefano Zampini PetscMPIInt psum; 511228143c3dSStefano Zampini PetscInt j; 511328143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 511428143c3dSStefano Zampini PetscInt plen; 511528143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 511628143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 511728143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 511828143c3dSStefano Zampini } 5119854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 512028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 512128143c3dSStefano Zampini PetscInt plen; 512228143c3dSStefano Zampini const PetscInt *is_array_idxs; 512328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 512428143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 512528143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 512628143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 512728143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 512828143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 512928143c3dSStefano Zampini } 513028143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 513128143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 513228143c3dSStefano Zampini } 513328143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 513428143c3dSStefano Zampini } 513528143c3dSStefano Zampini 5136e7931f94SStefano Zampini buf_size_idxs = 0; 5137e7931f94SStefano Zampini buf_size_vals = 0; 513828143c3dSStefano Zampini buf_size_idxs_is = 0; 5139e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5140e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5141e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 514228143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5143e7931f94SStefano Zampini } 5144785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5145785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 514695ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5147e7931f94SStefano Zampini 5148e7931f94SStefano Zampini /* get new tags for clean communications */ 5149e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5150e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 515128143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5152e7931f94SStefano Zampini 5153e7931f94SStefano Zampini /* allocate for requests */ 5154785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5155785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 515695ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5157785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5158785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 515995ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5160e7931f94SStefano Zampini 5161e7931f94SStefano Zampini /* communications */ 5162e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5163e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 516428143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5165e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5166e7931f94SStefano Zampini source_dest = onodes[i]; 5167e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5168e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5169e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5170e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 517128143c3dSStefano Zampini if (nis) { 517257de7509SStefano Zampini source_dest = onodes_is[i]; 517328143c3dSStefano 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); 517428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 517528143c3dSStefano Zampini } 5176e7931f94SStefano Zampini } 5177e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5178e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5179e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5180e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 518128143c3dSStefano Zampini if (nis) { 518228143c3dSStefano 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); 518328143c3dSStefano Zampini } 5184e7931f94SStefano Zampini } 5185e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5186e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5187e7931f94SStefano Zampini 5188e7931f94SStefano Zampini /* assemble new l2g map */ 5189e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5190e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51919d30be91SStefano Zampini new_local_rows = 0; 5192e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51939d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5194e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5195e7931f94SStefano Zampini } 51969d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5197e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51989d30be91SStefano Zampini new_local_rows = 0; 5199e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 52009d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 52019d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5202e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5203e7931f94SStefano Zampini } 52049d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 52059d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5206e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5207e7931f94SStefano Zampini 5208e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5209e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5210e7931f94SStefano 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) */ 5211e7931f94SStefano Zampini if (n_recvs) { 521228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5213e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5214e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5215e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5216e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5217e7931f94SStefano Zampini break; 5218e7931f94SStefano Zampini } 5219e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5220e7931f94SStefano Zampini } 5221e7931f94SStefano Zampini switch (new_local_type_private) { 522228143c3dSStefano Zampini case MATDENSE_PRIVATE: 522328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5224e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5225e7931f94SStefano Zampini bs = 1; 522628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 522728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 522828143c3dSStefano Zampini bs = 1; 522928143c3dSStefano Zampini } 5230e7931f94SStefano Zampini break; 5231e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5232e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5233e7931f94SStefano Zampini bs = 1; 5234e7931f94SStefano Zampini break; 5235e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5236e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5237e7931f94SStefano Zampini break; 5238e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5239e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5240e7931f94SStefano Zampini break; 5241e7931f94SStefano Zampini default: 52429d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5243e7931f94SStefano Zampini break; 5244e7931f94SStefano Zampini } 524528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 524628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 524728143c3dSStefano Zampini bs = 1; 5248e7931f94SStefano Zampini } 5249e7931f94SStefano Zampini 525070cf5478SStefano Zampini /* create MATIS object if needed */ 525157de7509SStefano Zampini if (!reuse) { 5252e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5253e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 525470cf5478SStefano Zampini } else { 525570cf5478SStefano Zampini /* it also destroys the local matrices */ 525657de7509SStefano Zampini if (*mat_n) { 525770cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 525857de7509SStefano Zampini } else { /* this is a fake object */ 525957de7509SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 526057de7509SStefano Zampini } 526170cf5478SStefano Zampini } 526270cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5263e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 52649d30be91SStefano Zampini 52659d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 52669d30be91SStefano Zampini 52679d30be91SStefano Zampini /* Global to local map of received indices */ 52689d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 52699d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 52709d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 52719d30be91SStefano Zampini 52729d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 52739d30be91SStefano Zampini buf_size_idxs = 0; 52749d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 52759d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 52769d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 52779d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 52789d30be91SStefano Zampini } 52799d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 52809d30be91SStefano Zampini 52819d30be91SStefano Zampini /* set preallocation */ 52829d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 52839d30be91SStefano Zampini if (!newisdense) { 52849d30be91SStefano Zampini PetscInt *new_local_nnz=0; 52859d30be91SStefano Zampini 52869d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 52879d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 52889d30be91SStefano Zampini if (n_recvs) { 52899d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 52909d30be91SStefano Zampini } 52919d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 52929d30be91SStefano Zampini PetscInt j; 52939d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 52949d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 52959d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 52969d30be91SStefano Zampini } 52979d30be91SStefano Zampini } else { 52989d30be91SStefano Zampini /* TODO */ 52999d30be91SStefano Zampini } 53009d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 53019d30be91SStefano Zampini } 53029d30be91SStefano Zampini if (new_local_nnz) { 53039d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 53049d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 53059d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 53069d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 53079d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 53089d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 53099d30be91SStefano Zampini } else { 53109d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 53119d30be91SStefano Zampini } 53129d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 53139d30be91SStefano Zampini } else { 53149d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 53159d30be91SStefano Zampini } 5316e7931f94SStefano Zampini 5317e7931f94SStefano Zampini /* set values */ 5318e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 53199d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5320e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5321e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5322e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 53239d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5324e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5325e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5326e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 532728143c3dSStefano Zampini } else { 532828143c3dSStefano Zampini /* TODO */ 5329e7931f94SStefano Zampini } 5330e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5331e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5332e7931f94SStefano Zampini } 5333e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5334e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 533570cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 533670cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53379d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 53389d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5339e7931f94SStefano Zampini 5340dfd14d43SStefano Zampini #if 0 534128143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5342e7931f94SStefano Zampini Vec lvec,rvec; 5343e7931f94SStefano Zampini PetscReal infty_error; 5344e7931f94SStefano Zampini 53452a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5346e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5347e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5348e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 534970cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5350e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5351e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5352e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5353e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5354e7931f94SStefano Zampini } 535528143c3dSStefano Zampini #endif 5356e7931f94SStefano Zampini 535728143c3dSStefano Zampini /* assemble new additional is (if any) */ 535828143c3dSStefano Zampini if (nis) { 535928143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 536028143c3dSStefano Zampini 536128143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5362854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 536328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 536428143c3dSStefano Zampini psum = 0; 536528143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 536628143c3dSStefano Zampini for (j=0;j<nis;j++) { 536728143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 536828143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 536928143c3dSStefano Zampini psum += plen; 537028143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 537128143c3dSStefano Zampini } 537228143c3dSStefano Zampini } 5373854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5374854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 537528143c3dSStefano Zampini for (i=1;i<nis;i++) { 537628143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 537728143c3dSStefano Zampini } 537828143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 537928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 538028143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 538128143c3dSStefano Zampini for (j=0;j<nis;j++) { 538228143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 538328143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 538428143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 538528143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 538628143c3dSStefano Zampini } 538728143c3dSStefano Zampini } 538828143c3dSStefano Zampini for (i=0;i<nis;i++) { 538928143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 539028143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 539128143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 539228143c3dSStefano Zampini } 539328143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 539428143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 539528143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 539628143c3dSStefano Zampini } 5397e7931f94SStefano Zampini /* free workspace */ 539828143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5399e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5400e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5401e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5402e7931f94SStefano Zampini if (isdense) { 5403e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5404e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5405e7931f94SStefano Zampini } else { 5406e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5407e7931f94SStefano Zampini } 540828143c3dSStefano Zampini if (nis) { 540928143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 541028143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 541128143c3dSStefano Zampini } 5412e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5413e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 541428143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5415e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5416e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 541728143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5418e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5419e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5420e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5421e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5422e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 542328143c3dSStefano Zampini if (nis) { 542428143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 542528143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 542628143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 542728143c3dSStefano Zampini } 542828143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 542928143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 543028143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 543128143c3dSStefano Zampini for (i=0;i<nis;i++) { 543228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 543328143c3dSStefano Zampini } 543453a05cb3SStefano Zampini *mat_n = NULL; 543528143c3dSStefano Zampini } 5436e7931f94SStefano Zampini PetscFunctionReturn(0); 5437e7931f94SStefano Zampini } 5438a57a6d2fSStefano Zampini 543912edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5440af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 544112edc857SStefano Zampini 5442c8587f34SStefano Zampini #undef __FUNCT__ 5443c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5444c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5445c8587f34SStefano Zampini { 5446c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5447c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 544820a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 54499881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 545020a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 54516e683305SStefano Zampini IS coarse_is,*isarray; 54526e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 545330368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5454f9eb5b7dSStefano Zampini PC pc_temp; 5455c8587f34SStefano Zampini PCType coarse_pc_type; 5456c8587f34SStefano Zampini KSPType coarse_ksp_type; 5457f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 54584f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 54596e683305SStefano Zampini Mat t_coarse_mat_is; 546057de7509SStefano Zampini PetscInt ncoarse; 546168457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 546222bc73bbSStefano Zampini PetscScalar *array; 546357de7509SStefano Zampini MatReuse coarse_mat_reuse; 546457de7509SStefano Zampini PetscBool restr, full_restr, have_void; 54659881197aSStefano Zampini PetscErrorCode ierr; 5466fdc09c96SStefano Zampini 5467c8587f34SStefano Zampini PetscFunctionBegin; 5468c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 546968457ee5SStefano 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 */ 5470fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 54715a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5472fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5473f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5474f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5475f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5476fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 547751bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 547851bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5479dc4bcba2SStefano Zampini PC pc; 5480dc4bcba2SStefano Zampini PetscBool isbddc; 5481dc4bcba2SStefano Zampini 5482dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5483dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5484dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5485dc4bcba2SStefano Zampini if (isbddc) { 548663c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 548763c961adSStefano Zampini } else { 5488727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 548963c961adSStefano Zampini } 5490fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5491fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5492fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5493f4ddd8eeSStefano Zampini } 5494fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5495fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5496f4ddd8eeSStefano Zampini } 549770cf5478SStefano Zampini /* reset any subassembling information */ 549857de7509SStefano Zampini if (!coarse_reuse || pcbddc->recompute_topography) { 549970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 550057de7509SStefano Zampini } 55016e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5502fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5503f4ddd8eeSStefano Zampini } 550457de7509SStefano Zampini /* assemble coarse matrix */ 550557de7509SStefano Zampini if (coarse_reuse && pcbddc->coarse_ksp) { 550657de7509SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 550757de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 550857de7509SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 550918a45a71SStefano Zampini } else { 551057de7509SStefano Zampini coarse_mat = NULL; 551157de7509SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 55126e683305SStefano Zampini } 5513e7931f94SStefano Zampini 5514abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5515abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5516abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5517abbbba34SStefano Zampini 5518abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 551922bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 552022bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 552122bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 552222bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5523e176bc59SStefano 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); 55246e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 55256e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55266e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5527abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5528abbbba34SStefano Zampini 552957de7509SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 553057de7509SStefano Zampini im_active = !!(pcis->n); 553157de7509SStefano Zampini ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 553257de7509SStefano Zampini 553357de7509SStefano Zampini /* determine number of process partecipating to coarse solver and compute subassembling pattern */ 553457de7509SStefano Zampini /* restr : whether if we want to exclude senders (which are not receivers) from the subassembling pattern */ 553557de7509SStefano Zampini /* full_restr : just use the receivers from the subassembling pattern */ 553657de7509SStefano Zampini coarse_mat_is = NULL; 553757de7509SStefano Zampini multilevel_allowed = PETSC_FALSE; 553857de7509SStefano Zampini multilevel_requested = PETSC_FALSE; 553957de7509SStefano Zampini full_restr = PETSC_TRUE; 554057de7509SStefano Zampini pcbddc->coarse_eqs_per_proc = PetscMin(pcbddc->coarse_size,pcbddc->coarse_eqs_per_proc); 554157de7509SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE; 554257de7509SStefano Zampini if (multilevel_requested) { 554357de7509SStefano Zampini ncoarse = active_procs/pcbddc->coarsening_ratio; 554457de7509SStefano Zampini restr = PETSC_FALSE; 554557de7509SStefano Zampini full_restr = PETSC_FALSE; 554657de7509SStefano Zampini } else { 554757de7509SStefano Zampini ncoarse = pcbddc->coarse_size/pcbddc->coarse_eqs_per_proc; 554857de7509SStefano Zampini restr = PETSC_TRUE; 554957de7509SStefano Zampini full_restr = PETSC_TRUE; 555057de7509SStefano Zampini } 555157de7509SStefano Zampini ncoarse = PetscMax(1,ncoarse); 555257de7509SStefano Zampini if (!pcbddc->coarse_subassembling) { 555357de7509SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);CHKERRQ(ierr); 555457de7509SStefano Zampini } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */ 555557de7509SStefano Zampini PetscInt psum; 555657de7509SStefano Zampini PetscMPIInt size; 555757de7509SStefano Zampini if (pcbddc->coarse_ksp) psum = 1; 555857de7509SStefano Zampini else psum = 0; 555957de7509SStefano Zampini ierr = MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 556057de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);CHKERRQ(ierr); 556157de7509SStefano Zampini if (ncoarse < size) have_void = PETSC_TRUE; 556257de7509SStefano Zampini } 556357de7509SStefano Zampini /* determine if we can go multilevel */ 556457de7509SStefano Zampini if (multilevel_requested) { 556557de7509SStefano Zampini if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */ 556657de7509SStefano Zampini else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */ 556757de7509SStefano Zampini } 556857de7509SStefano Zampini if (multilevel_allowed && have_void) restr = PETSC_TRUE; 556957de7509SStefano Zampini 55706e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 557130368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 55726e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 55736e683305SStefano Zampini const PetscInt *idxs; 55746e683305SStefano Zampini ISLocalToGlobalMapping tmap; 55756e683305SStefano Zampini 55766e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 55770be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 55786e683305SStefano Zampini /* allocate space for temporary storage */ 5579854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5580854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 55816e683305SStefano Zampini /* allocate for IS array */ 55826e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 55836e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 558430368db7SStefano Zampini nisvert = 0; 558530368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 558630368db7SStefano Zampini nisvert = 1; 558730368db7SStefano Zampini } 558830368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5589854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 55906e683305SStefano Zampini /* dofs splitting */ 55916e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 55926e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 55936e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 55946e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55956e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55966e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55976e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 559830368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 55996e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 56006e683305SStefano Zampini } 56016e683305SStefano Zampini /* neumann boundaries */ 56026e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 56036e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 56046e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 56056e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 56066e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 56076e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 56086e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 560930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 56106e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 56116e683305SStefano Zampini } 561230368db7SStefano Zampini /* primal vertices (benign) */ 561330368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 561430368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 561530368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 561630368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 561730368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 561830368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 561930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 5620b9b0e38cSStefano Zampini /* ierr = ISView(isarray[nis-1],0);CHKERRQ(ierr); */ 562130368db7SStefano Zampini } 56226e683305SStefano Zampini /* free memory */ 56236e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 56246e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 56256e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 56266e683305SStefano Zampini } else { 56276e683305SStefano Zampini nis = 0; 56286e683305SStefano Zampini nisdofs = 0; 56296e683305SStefano Zampini nisneu = 0; 563030368db7SStefano Zampini nisvert = 0; 56316e683305SStefano Zampini isarray = NULL; 56326e683305SStefano Zampini } 56336e683305SStefano Zampini /* destroy no longer needed map */ 56346e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 56356e683305SStefano Zampini 563657de7509SStefano Zampini /* subassemble */ 563757de7509SStefano Zampini if (multilevel_allowed) { 563857de7509SStefano Zampini PetscBool reuse,reuser; 563957de7509SStefano Zampini if (coarse_mat) reuse = PETSC_TRUE; 564057de7509SStefano Zampini else reuse = PETSC_FALSE; 564157de7509SStefano Zampini ierr = MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 564257de7509SStefano Zampini if (reuser) { 564357de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray);CHKERRQ(ierr); 564474e2c79eSStefano Zampini } else { 564557de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 564674e2c79eSStefano Zampini } 564774e2c79eSStefano Zampini } else { 564857de7509SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 56496e683305SStefano Zampini } 565057de7509SStefano Zampini if (coarse_mat_is || coarse_mat) { 565157de7509SStefano Zampini PetscMPIInt size; 565257de7509SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)coarse_mat_is),&size); 565357de7509SStefano Zampini if (!multilevel_allowed) { 565457de7509SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 56556e683305SStefano Zampini } else { 565657de7509SStefano Zampini Mat A; 5657779c1cceSStefano Zampini 565857de7509SStefano Zampini /* if this matrix is present, it means we are not reusing the coarse matrix */ 565957de7509SStefano Zampini if (coarse_mat_is) { 566057de7509SStefano Zampini if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen"); 566157de7509SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 566257de7509SStefano Zampini coarse_mat = coarse_mat_is; 566357de7509SStefano Zampini } 566457de7509SStefano Zampini /* be sure we don't have MatSeqDENSE as local mat */ 566557de7509SStefano Zampini ierr = MatISGetLocalMat(coarse_mat,&A);CHKERRQ(ierr); 566657de7509SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INPLACE_MATRIX,&A);CHKERRQ(ierr); 5667779c1cceSStefano Zampini } 5668779c1cceSStefano Zampini } 566957de7509SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 567057de7509SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 56716e683305SStefano Zampini 56726e683305SStefano Zampini /* create local to global scatters for coarse problem */ 567368457ee5SStefano Zampini if (compute_vecs) { 56746e683305SStefano Zampini PetscInt lrows; 56756e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 567657de7509SStefano Zampini if (coarse_mat) { 567757de7509SStefano Zampini ierr = MatGetLocalSize(coarse_mat,&lrows,NULL);CHKERRQ(ierr); 56786e683305SStefano Zampini } else { 56796e683305SStefano Zampini lrows = 0; 56806e683305SStefano Zampini } 56816e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 56826e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 56836e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 56846e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56856e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56866e683305SStefano Zampini } 56876e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 5688c8587f34SStefano Zampini 5689f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5690f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5691f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5692f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5693f9eb5b7dSStefano Zampini } else { 5694f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5695f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5696c8587f34SStefano Zampini } 5697c8587f34SStefano Zampini 56986e683305SStefano Zampini /* print some info if requested */ 56996e683305SStefano Zampini if (pcbddc->dbg_flag) { 57006e683305SStefano Zampini if (!multilevel_allowed) { 57016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57026e683305SStefano Zampini if (multilevel_requested) { 57036e683305SStefano 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); 57046e683305SStefano Zampini } else if (pcbddc->max_levels) { 57056e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 57066e683305SStefano Zampini } 57076e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 57086e683305SStefano Zampini } 57096e683305SStefano Zampini } 57106e683305SStefano Zampini 5711f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 571257de7509SStefano Zampini if (coarse_mat) { 57136a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 57146e683305SStefano Zampini if (pcbddc->dbg_flag) { 571557de7509SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat)); 57166e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57176e683305SStefano Zampini } 5718f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5719312be037SStefano Zampini char prefix[256],str_level[16]; 5720e604994aSStefano Zampini size_t len; 572157de7509SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5722422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5723c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5724f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 572557de7509SStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5726c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 57276e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5728c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5729c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5730e604994aSStefano Zampini /* prefix */ 5731e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5732e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5733e604994aSStefano Zampini if (!pcbddc->current_level) { 5734e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5735e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5736c8587f34SStefano Zampini } else { 5737e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5738312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5739312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 574034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5741312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5742e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5743e604994aSStefano Zampini } 5744e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 57453e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 57463e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 57473e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 57483e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5749f9eb5b7dSStefano Zampini /* allow user customization */ 5750f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 57513e3c6dadSStefano Zampini } 57523e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 575351bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 57543e3c6dadSStefano Zampini if (nisdofs) { 57553e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 57563e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 57573e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 57583e3c6dadSStefano Zampini } 57593e3c6dadSStefano Zampini } 57603e3c6dadSStefano Zampini if (nisneu) { 57613e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 57623e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5763312be037SStefano Zampini } 576430368db7SStefano Zampini if (nisvert) { 576530368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 576630368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 576730368db7SStefano Zampini } 5768f9eb5b7dSStefano Zampini 5769f9eb5b7dSStefano Zampini /* get some info after set from options */ 5770f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5771f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 57724f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 57736e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5774f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5775f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5776f9eb5b7dSStefano Zampini } 577739f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 57784f3a063dSStefano Zampini if (isredundant) { 57794f3a063dSStefano Zampini KSP inner_ksp; 57804f3a063dSStefano Zampini PC inner_pc; 57814f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 57824f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 57834f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 57844f3a063dSStefano Zampini } 5785f9eb5b7dSStefano Zampini 578657de7509SStefano Zampini /* parameters which miss an API */ 578757de7509SStefano Zampini if (isbddc) { 5788720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5789720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 579057de7509SStefano Zampini pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc; 5791720d30f9SStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_saddle_point; 5792720d30f9SStefano Zampini pcbddc_coarse->benign_compute_nonetflux = pcbddc->benign_compute_nonetflux; 5793720d30f9SStefano Zampini if (pcbddc_coarse->benign_compute_nonetflux) { 5794720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 5795720d30f9SStefano Zampini } 5796d4d8cf7bSStefano Zampini } 57979881197aSStefano Zampini 57983301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 57995a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 58003301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 58013301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 58023301b35fSStefano Zampini } 58033301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 58043301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 58053301b35fSStefano Zampini } 58063301b35fSStefano Zampini if (pc->pmat->spd_set) { 58073301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 58083301b35fSStefano Zampini } 58096e683305SStefano Zampini /* set operators */ 58105f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 58116e683305SStefano Zampini if (pcbddc->dbg_flag) { 58126e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 58136e683305SStefano Zampini } 58146e683305SStefano Zampini } 58156e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5816b1ecc7b1SStefano Zampini #if 0 5817b9b85e73SStefano Zampini { 5818b9b85e73SStefano Zampini PetscViewer viewer; 5819b9b85e73SStefano Zampini char filename[256]; 5820b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5821b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 58226a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5823b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5824f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 5825b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5826b9b85e73SStefano Zampini } 5827b9b85e73SStefano Zampini #endif 5828f9eb5b7dSStefano Zampini 5829b8ffe317SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5830298c0119SStefano Zampini #if 0 5831b8ffe317SStefano Zampini if (pcbddc->NullSpace) { 5832b8ffe317SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 583398a51de6SStefano Zampini } 5834298c0119SStefano Zampini #endif 5835b0f5fe93SStefano Zampini /* hack */ 583698a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 583798a51de6SStefano Zampini Vec crhs,csol; 583804708bb6SStefano Zampini 5839f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5840f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5841f347579bSStefano Zampini if (!csol) { 58422a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5843f9eb5b7dSStefano Zampini } 5844f347579bSStefano Zampini if (!crhs) { 58452a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5846f347579bSStefano Zampini } 5847b0f5fe93SStefano Zampini } 5848b0f5fe93SStefano Zampini 5849b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5850b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5851b0f5fe93SStefano Zampini 5852b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 58534f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58544f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 58554f1b2e48SStefano Zampini } 5856b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5857b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5858b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5859b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5860b0f5fe93SStefano Zampini if (coarse_mat) { 5861b0f5fe93SStefano Zampini Vec nullv; 5862b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5863b0f5fe93SStefano Zampini PetscInt nl; 5864b0f5fe93SStefano Zampini 5865b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5866b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5867b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5868b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5869b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5870b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5871b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5872b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5873b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5874b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5875b0f5fe93SStefano Zampini } 5876b0f5fe93SStefano Zampini } 5877b0f5fe93SStefano Zampini 5878b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5879b0f5fe93SStefano Zampini PetscBool ispreonly; 5880b0f5fe93SStefano Zampini 5881b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5882b0f5fe93SStefano Zampini PetscBool isnull; 5883b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5884d4d8cf7bSStefano Zampini if (0) { 588530368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5886b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5887b0f5fe93SStefano Zampini } else { 5888b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5889b0f5fe93SStefano Zampini } 5890b0f5fe93SStefano Zampini } else { 5891b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5892b0f5fe93SStefano Zampini } 5893b0f5fe93SStefano Zampini } 5894b0f5fe93SStefano Zampini /* setup coarse ksp */ 5895b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5896cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5897cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 58986e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5899c8587f34SStefano Zampini KSP check_ksp; 59002b510759SStefano Zampini KSPType check_ksp_type; 5901c8587f34SStefano Zampini PC check_pc; 59026e683305SStefano Zampini Vec check_vec,coarse_vec; 59036a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 59042b510759SStefano Zampini PetscInt its; 59056e683305SStefano Zampini PetscBool compute_eigs; 59066e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 59076e683305SStefano Zampini PetscInt neigs; 59088e185a42SStefano Zampini const char *prefix; 5909c8587f34SStefano Zampini 59102b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 59116e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5912422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 591323ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5914f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 59152b510759SStefano Zampini if (ispreonly) { 59162b510759SStefano Zampini check_ksp_type = KSPPREONLY; 59176e683305SStefano Zampini compute_eigs = PETSC_FALSE; 59182b510759SStefano Zampini } else { 5919cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 59206e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5921c8587f34SStefano Zampini } 5922c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 59236e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 59246e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 59256e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5926a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5927a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5928a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5929a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5930c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5931c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5932c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5933c8587f34SStefano Zampini /* create random vec */ 59346e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 59356e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5936c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5937c8587f34SStefano Zampini if (CoarseNullSpace) { 5938c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5939c8587f34SStefano Zampini } 59406e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5941c8587f34SStefano Zampini /* solve coarse problem */ 59426e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5943c8587f34SStefano Zampini if (CoarseNullSpace) { 59446e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5945c8587f34SStefano Zampini } 5946cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 59476e683305SStefano Zampini if (compute_eigs) { 5948854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5949854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 59506e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 59516e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 59526e683305SStefano Zampini lambda_min = eigs_r[0]; 59536e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 59546e683305SStefano Zampini if (lambda_max>lambda_min) { 5955cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5956cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5957cbcc2c2aSStefano Zampini } 5958c8587f34SStefano Zampini } 5959c8587f34SStefano Zampini } 5960cbcc2c2aSStefano Zampini 5961c8587f34SStefano Zampini /* check coarse problem residual error */ 59626e683305SStefano Zampini if (pcbddc->dbg_flag) { 59636e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 59646e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59656e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5966c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 59676e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 59686e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5969c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5970779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 59716e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 59726e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 59736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 59746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5975b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5976b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5977b0f5fe93SStefano Zampini } 59786e683305SStefano Zampini if (compute_eigs) { 59796e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5980deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5981c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 59826e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 59836e683305SStefano 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); 59846e683305SStefano Zampini for (i=0;i<neigs;i++) { 59856e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5986c8587f34SStefano Zampini } 59876e683305SStefano Zampini } 59886e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 59896e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59906e683305SStefano Zampini } 5991c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 59926e683305SStefano Zampini if (compute_eigs) { 59936e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 59946e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5995c8587f34SStefano Zampini } 59966e683305SStefano Zampini } 59976e683305SStefano Zampini } 5998cbcc2c2aSStefano Zampini /* print additional info */ 5999cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 60006e683305SStefano Zampini /* waits until all processes reaches this point */ 60016e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6002cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6003cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6004cbcc2c2aSStefano Zampini } 6005cbcc2c2aSStefano Zampini 60062b510759SStefano Zampini /* free memory */ 6007c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6008fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6009c8587f34SStefano Zampini PetscFunctionReturn(0); 6010c8587f34SStefano Zampini } 6011674ae819SStefano Zampini 6012f34684f1SStefano Zampini #undef __FUNCT__ 6013f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6014f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6015f34684f1SStefano Zampini { 6016f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6017f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6018f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6019dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6020dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 602173be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6022dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6023f34684f1SStefano Zampini PetscErrorCode ierr; 6024f34684f1SStefano Zampini 6025f34684f1SStefano Zampini PetscFunctionBegin; 6026f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 60276c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6028dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 60293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6030dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6031dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6032dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6033dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6034dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6035dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60366c4ed002SBarry 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); 6037dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6038dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6039dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6040dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6041dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6042f34684f1SStefano Zampini 6043f34684f1SStefano Zampini /* check numbering */ 6044f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6045019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6046dc456d91SStefano Zampini PetscInt i; 6047b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6048f34684f1SStefano Zampini 6049f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6050f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6051f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 60521575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6053019a44ceSStefano Zampini /* counter */ 6054019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6055019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6056019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6057019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6058019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6059019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6060f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6061f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6062727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6063f34684f1SStefano Zampini } 6064f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6065f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6066f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6067e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6068e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6069e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6070e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6071f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6072019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6073f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6074019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 607575c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 607675c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6077b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6078019a44ceSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %d owned by a %d processes instead of %d!\n",PetscGlobalRank,i,owned,neigh);CHKERRQ(ierr); 6079f34684f1SStefano Zampini } 6080f34684f1SStefano Zampini } 6081019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6082b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6083f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6084f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6085f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6086f34684f1SStefano Zampini } 6087f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6088f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6089e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6090e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6091f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6092f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6093b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6094ca8b9ea9SStefano Zampini PetscInt *gidxs; 6095ca8b9ea9SStefano Zampini 6096ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 60973bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6098f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6099f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6100f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6101f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 61024bc2dc4bSStefano 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); 6103f34684f1SStefano Zampini } 6104f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6105ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6106f34684f1SStefano Zampini } 6107f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 61081575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6109302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6110f34684f1SStefano Zampini } 61118bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6112f34684f1SStefano Zampini /* get back data */ 6113f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6114f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6115674ae819SStefano Zampini PetscFunctionReturn(0); 6116674ae819SStefano Zampini } 6117674ae819SStefano Zampini 6118e456f2a8SStefano Zampini #undef __FUNCT__ 6119e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6120a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6121e456f2a8SStefano Zampini { 6122e456f2a8SStefano Zampini IS localis_t; 6123a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6124e456f2a8SStefano Zampini PetscScalar *vals; 6125e456f2a8SStefano Zampini PetscErrorCode ierr; 6126e456f2a8SStefano Zampini 6127e456f2a8SStefano Zampini PetscFunctionBegin; 6128a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6129e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6130854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6131e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6132e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6133a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6134a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61351035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6136a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61371035eff8SStefano Zampini } 6138a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6139e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6140e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6141a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6142a7dc3881SStefano Zampini /* now compute set in local ordering */ 6143a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6144a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6145a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6146a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6147a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6148ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6149e456f2a8SStefano Zampini lsize++; 6150e456f2a8SStefano Zampini } 6151e456f2a8SStefano Zampini } 6152854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6153a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6154ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6155e456f2a8SStefano Zampini idxs[lsize++] = i; 6156e456f2a8SStefano Zampini } 6157e456f2a8SStefano Zampini } 6158a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6159a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6160e456f2a8SStefano Zampini *localis = localis_t; 6161e456f2a8SStefano Zampini PetscFunctionReturn(0); 6162e456f2a8SStefano Zampini } 6163906d46d4SStefano Zampini 6164b96c3477SStefano Zampini #undef __FUNCT__ 6165b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 616608122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6167b96c3477SStefano Zampini { 6168a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6169b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6170b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6171a64f4aa4SStefano Zampini Mat S_j; 6172b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6173b96c3477SStefano Zampini PetscBool free_used_adj; 6174b96c3477SStefano Zampini PetscErrorCode ierr; 6175b96c3477SStefano Zampini 6176b96c3477SStefano Zampini PetscFunctionBegin; 6177b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6178b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 617908122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6180b96c3477SStefano Zampini used_xadj = NULL; 6181b96c3477SStefano Zampini used_adjncy = NULL; 6182b96c3477SStefano Zampini } else { 618308122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 618408122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 618508122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 618608122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6187b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6188b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6189b96c3477SStefano Zampini } else { 61902fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6191b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6192b96c3477SStefano Zampini PetscInt nvtxs; 6193b96c3477SStefano Zampini 61942fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 61952fffb893SStefano Zampini if (flg_row) { 6196b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6197b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6198b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6199b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62002fffb893SStefano Zampini } else { 62012fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62022fffb893SStefano Zampini used_xadj = NULL; 62032fffb893SStefano Zampini used_adjncy = NULL; 62042fffb893SStefano Zampini } 62052fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6206b96c3477SStefano Zampini } 6207b96c3477SStefano Zampini } 6208d5574798SStefano Zampini 6209d5574798SStefano Zampini /* setup sub_schurs data */ 6210a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6211df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6212df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6213a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 621491af6908SStefano 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); 6215a64f4aa4SStefano Zampini } else { 62166816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 6217b7ab4a40SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE;; 6218a3df083aSStefano Zampini PetscInt benign_n; 621972b8c272SStefano Zampini Mat change = NULL; 62209d54b7f4SStefano Zampini Vec scaling = NULL; 622172b8c272SStefano Zampini IS change_primal = NULL; 6222a3df083aSStefano Zampini 62235feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62245feab87aSStefano Zampini PetscInt n_vertices; 62255feab87aSStefano Zampini 62265feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62272034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62285feab87aSStefano Zampini } 622904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 623004708bb6SStefano Zampini if (!isseqaij) { 623104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 623204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 623304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 623404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 623504708bb6SStefano Zampini } else { 6236511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 623704708bb6SStefano Zampini } 623804708bb6SStefano Zampini } 6239a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6240a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6241ca92afb2SStefano Zampini } else { 6242a3df083aSStefano Zampini benign_n = 0; 6243ca92afb2SStefano Zampini } 6244b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6245b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6246b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 624772b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 6248b7ab4a40SStefano Zampini PetscBool have_loc_change = !!(sub_schurs->change); 6249b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6250b7ab4a40SStefano Zampini need_change = !need_change; 6251b7ab4a40SStefano Zampini } 6252b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6253b7ab4a40SStefano 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 */ 6254b7ab4a40SStefano Zampini if (need_change) { 625588c03ad3SStefano Zampini PC_IS *pcisf; 625688c03ad3SStefano Zampini PC_BDDC *pcbddcf; 625788c03ad3SStefano Zampini PC pcf; 625888c03ad3SStefano Zampini 625988c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 626088c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 626188c03ad3SStefano Zampini } 626288c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 626388c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 626488c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 626588c03ad3SStefano Zampini /* hacks */ 626688c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 626772b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 626872b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 626972b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 627072b8c272SStefano Zampini pcisf->n = pcis->n; 627172b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 627288c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 627388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 627488c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 627588c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 627688c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 627788c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 627872b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 627988c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 628088c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 628172b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 628272b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 628372b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 628472b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 628572b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 628688c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 628772b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 628888c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 628988c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 629088c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 629188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 629288c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 629388c03ad3SStefano Zampini pcf->ops->destroy = NULL; 629488c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 629588c03ad3SStefano Zampini } 62969d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 629791af6908SStefano 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); 629872b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 629972b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6300ca92afb2SStefano Zampini } 6301d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6302b96c3477SStefano Zampini 6303b96c3477SStefano Zampini /* free adjacency */ 6304b96c3477SStefano Zampini if (free_used_adj) { 6305b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6306b96c3477SStefano Zampini } 6307b96c3477SStefano Zampini PetscFunctionReturn(0); 6308b96c3477SStefano Zampini } 6309b96c3477SStefano Zampini 6310b96c3477SStefano Zampini #undef __FUNCT__ 6311b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 631208122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6313b96c3477SStefano Zampini { 6314b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6315b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6316b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6317b96c3477SStefano Zampini PCBDDCGraph graph; 6318b96c3477SStefano Zampini PetscErrorCode ierr; 6319b96c3477SStefano Zampini 6320b96c3477SStefano Zampini PetscFunctionBegin; 6321b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 632208122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63233301b35fSStefano Zampini IS verticesIS,verticescomm; 63243301b35fSStefano Zampini PetscInt vsize,*idxs; 6325b96c3477SStefano Zampini 6326b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63273301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63283301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63293301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63303301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63313301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6332b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63337fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6334441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63353301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6336b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6337b96c3477SStefano Zampini /* 6338b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6339b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6340b96c3477SStefano Zampini } 6341b96c3477SStefano Zampini */ 6342b96c3477SStefano Zampini } else { 6343b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6344b96c3477SStefano Zampini } 6345b96c3477SStefano Zampini 6346b96c3477SStefano Zampini /* sub_schurs init */ 6347a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6348a64f4aa4SStefano Zampini 6349b96c3477SStefano Zampini /* free graph struct */ 635008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6351b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6352b96c3477SStefano Zampini } 6353b96c3477SStefano Zampini PetscFunctionReturn(0); 6354b96c3477SStefano Zampini } 6355fa34dd3eSStefano Zampini 6356fa34dd3eSStefano Zampini #undef __FUNCT__ 6357fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6358fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6359fa34dd3eSStefano Zampini { 6360fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6361fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6362fa34dd3eSStefano Zampini PetscErrorCode ierr; 6363fa34dd3eSStefano Zampini 6364fa34dd3eSStefano Zampini PetscFunctionBegin; 6365fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6366fa34dd3eSStefano Zampini IS zerodiag = NULL; 63674f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6368fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 63694f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 637075c01103SStefano Zampini PetscReal norm; 6371fa34dd3eSStefano Zampini PetscInt i; 6372fa34dd3eSStefano Zampini 6373fa34dd3eSStefano Zampini /* B0 and B0_B */ 6374fa34dd3eSStefano Zampini if (zerodiag) { 6375fa34dd3eSStefano Zampini IS dummy; 6376fa34dd3eSStefano Zampini 63774f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 63784f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6379fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6380fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6381fa34dd3eSStefano Zampini } 6382fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6383fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6384fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6385fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6386fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6387fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6388fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6389fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6390fa34dd3eSStefano Zampini /* S_j */ 6391fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6392fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6393fa34dd3eSStefano Zampini 6394fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6395fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6396fa34dd3eSStefano Zampini /* continuous in primal space */ 6397fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6398fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6399fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6400fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 64014f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 64024f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6403fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6404fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6405fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6406fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6407fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6408fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6409fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6410fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6411fa34dd3eSStefano Zampini 6412fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6413fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6414fa34dd3eSStefano Zampini /* local with Schur */ 6415fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6416fa34dd3eSStefano Zampini if (zerodiag) { 6417fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64184f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6419fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6420fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6421fa34dd3eSStefano Zampini } 6422fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6423fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6424fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6425fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6426fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6427fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6428fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6429fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6430fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6431fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6432fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6433fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6434fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6435fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6436fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6437fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6438fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6439fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6440fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6441fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6442fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6443fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6444fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6445fa34dd3eSStefano Zampini if (zerodiag) { 6446fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6447fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64484f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6449fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6450fa34dd3eSStefano Zampini } 6451fa34dd3eSStefano Zampini /* BDDC */ 6452fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6453fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6454fa34dd3eSStefano Zampini 6455fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6456fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6457fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6458fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64594f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64604f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6461fa34dd3eSStefano Zampini } 64624f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6463fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6464fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6465fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6466fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6467fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6468fa34dd3eSStefano Zampini } 6469fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6470fa34dd3eSStefano Zampini } 6471