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); 4414f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 4425408967cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %f (should be numerically 0.)",PetscAbsScalar(vals[0])); 4435408967cSStefano Zampini } 4445408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4455408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 446d12d3064SStefano Zampini 447d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 448d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 449d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 450d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 451d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 452d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 453d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 454d12d3064SStefano Zampini if (count[idxs[i]]) { 455d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 456d12d3064SStefano Zampini } 457d4d8cf7bSStefano Zampini } 458d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 459d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4605408967cSStefano Zampini } 461dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4625408967cSStefano Zampini 4635408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4645408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4665408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4674f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4685408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4694f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4704f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4714f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 4724f1b2e48SStefano Zampini } 4735408967cSStefano Zampini } 4745408967cSStefano Zampini PetscFunctionReturn(0); 4755408967cSStefano Zampini } 4765408967cSStefano Zampini 4775408967cSStefano Zampini #undef __FUNCT__ 478339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 479339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 480339f8db1SStefano Zampini { 481339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4824f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 483b0f5fe93SStefano Zampini PetscInt nz,n; 4844f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 485339f8db1SStefano Zampini PetscErrorCode ierr; 486339f8db1SStefano Zampini 487339f8db1SStefano Zampini PetscFunctionBegin; 4889f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4899f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 490a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 491a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 492a3df083aSStefano Zampini } 493a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 494a3df083aSStefano Zampini pcbddc->benign_n = 0; 4954f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4964f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 4974f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 4984f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 4994f1b2e48SStefano Zampini since the local Schur complements are SPD 5004f1b2e48SStefano Zampini */ 5014f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5024f1b2e48SStefano Zampini have_null = PETSC_TRUE; 50340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5044f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5054f1b2e48SStefano Zampini 5064f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5074f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5084f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5094f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 510ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 51140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 51240fa8d13SStefano Zampini if (!sorted) { 51340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 51440fa8d13SStefano Zampini } 51540fa8d13SStefano Zampini } else { 51640fa8d13SStefano Zampini pressures = NULL; 51740fa8d13SStefano Zampini } 51897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 51997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 52097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 521339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 522339f8db1SStefano Zampini if (!sorted) { 523339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 524339f8db1SStefano Zampini } 525339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5264f1b2e48SStefano Zampini if (!nz) { 5274f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5284f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 52940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 53040fa8d13SStefano Zampini } 5314f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5324f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5334f1b2e48SStefano Zampini zerodiag_subs = NULL; 5344f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5354f1b2e48SStefano Zampini if (has_null_pressures) { 5364f1b2e48SStefano Zampini IS *subs; 5374f1b2e48SStefano Zampini PetscInt nsubs,i; 5384f1b2e48SStefano Zampini 5394f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5404f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5414f1b2e48SStefano Zampini if (nsubs > 1) { 5424f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5444f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5454f1b2e48SStefano Zampini IS t_zerodiag_subs; 5464f1b2e48SStefano Zampini PetscInt nl; 5474f1b2e48SStefano Zampini 5484f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5494f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5504f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5514f1b2e48SStefano Zampini if (nl) { 5524f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5534f1b2e48SStefano Zampini 5544f1b2e48SStefano Zampini if (pressures) { 5554f1b2e48SStefano Zampini IS t_pressure_subs; 5564f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5574f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5584f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5594f1b2e48SStefano Zampini } 5604f1b2e48SStefano Zampini if (valid) { 5614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini pcbddc->benign_n++; 5634f1b2e48SStefano Zampini } else { 5644f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5654f1b2e48SStefano Zampini } 5664f1b2e48SStefano Zampini } 5674f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini } 5704f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5714f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5724f1b2e48SStefano Zampini if (pressures) { 5734f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5744f1b2e48SStefano Zampini } 5754f1b2e48SStefano Zampini if (valid) { 5764f1b2e48SStefano Zampini pcbddc->benign_n = 1; 577ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5784f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5794f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5804f1b2e48SStefano Zampini } 5814f1b2e48SStefano Zampini } 5824f1b2e48SStefano Zampini } 5834f1b2e48SStefano Zampini 5844f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 5854f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5864f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5874f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5884f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5894f1b2e48SStefano Zampini } 5904f1b2e48SStefano Zampini 5914f1b2e48SStefano Zampini /* final check for null pressures */ 5924f1b2e48SStefano Zampini if (zerodiag && pressures) { 5934f1b2e48SStefano Zampini PetscInt nz,np; 5944f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5954f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 5964f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 5974f1b2e48SStefano Zampini } 5984f1b2e48SStefano Zampini 5994f1b2e48SStefano Zampini if (recompute_zerodiag) { 6004f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6014f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6024f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6044f1b2e48SStefano Zampini } else { 6054f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6064f1b2e48SStefano Zampini 6074f1b2e48SStefano Zampini nzn = 0; 6084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6094f1b2e48SStefano Zampini PetscInt ns; 6104f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6114f1b2e48SStefano Zampini nzn += ns; 6124f1b2e48SStefano Zampini } 6134f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6144f1b2e48SStefano Zampini nzn = 0; 6154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6164f1b2e48SStefano Zampini PetscInt ns,*idxs; 6174f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6184f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6194f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6204f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini nzn += ns; 6224f1b2e48SStefano Zampini } 6234f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6244f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6254f1b2e48SStefano Zampini } 6264f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6274f1b2e48SStefano Zampini } 6284f1b2e48SStefano Zampini 629b3afcdbeSStefano Zampini /* no-net-flux */ 630b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 631b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 632b3afcdbeSStefano Zampini MatNullSpace near_null_space; 633b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 634b3afcdbeSStefano Zampini Vec quad_vec; 635b3afcdbeSStefano Zampini PetscScalar *pvals; 636b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 637b3afcdbeSStefano Zampini 638b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 639b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 640b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 641b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 642b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 643b3afcdbeSStefano Zampini 644b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 645b3afcdbeSStefano Zampini np = 0; 646b3afcdbeSStefano Zampini if (zerodiag) { 647b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 648b3afcdbeSStefano Zampini } else if (pressures) { 649b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 650b3afcdbeSStefano Zampini } 651b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 652b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 653b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 654b3afcdbeSStefano Zampini if (np) { 655b3afcdbeSStefano Zampini const PetscInt *pidxs; 656b3afcdbeSStefano Zampini 657b3afcdbeSStefano Zampini if (zerodiag) { 658b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 659b3afcdbeSStefano Zampini } else if (pressures) { 660b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 661b3afcdbeSStefano Zampini } 662b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 663b3afcdbeSStefano Zampini if (zerodiag) { 664b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 665b3afcdbeSStefano Zampini } else if (pressures) { 666b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 667b3afcdbeSStefano Zampini } 668b3afcdbeSStefano Zampini } 669b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 670b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 671b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 672b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 673b3afcdbeSStefano Zampini 674b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 675b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 676b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = -1; 677b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 678b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 679b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 680b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 681b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 682b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 683b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) { 684b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 685b3afcdbeSStefano Zampini pvals[i] = 0.; 686b3afcdbeSStefano Zampini } 687b3afcdbeSStefano Zampini } 688b3afcdbeSStefano Zampini ierr = PetscMalloc1(matis->sf_nleaves,&dummyins);CHKERRQ(ierr); 689b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) dummyins[i] = i; 690b3afcdbeSStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf_nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 691b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 692b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 693b3afcdbeSStefano Zampini 694b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 695b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 696b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 697b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 698b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 699b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 700b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 701b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 702b3afcdbeSStefano Zampini } 703b3afcdbeSStefano Zampini 704b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7054f1b2e48SStefano Zampini if (has_null_pressures) { 7064f1b2e48SStefano Zampini IS zerodiagc; 7074f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7084f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7094f1b2e48SStefano Zampini 7104f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 711339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 712339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 713339f8db1SStefano Zampini /* local change of basis for pressures */ 714339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 71597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 716339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 717339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 718339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7194f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7204f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7214f1b2e48SStefano Zampini PetscInt nzs,j; 7224f1b2e48SStefano Zampini 7234f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7244f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7254f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7264f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7274f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7284f1b2e48SStefano Zampini } 729339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 730339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 731339f8db1SStefano Zampini /* set identity on velocities */ 732339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 733339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 734339f8db1SStefano Zampini } 7354f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7364f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7379f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7384f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 739339f8db1SStefano Zampini /* set change on pressures */ 7404f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7414f1b2e48SStefano Zampini PetscScalar *array; 7424f1b2e48SStefano Zampini PetscInt nzs; 7434f1b2e48SStefano Zampini 7444f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7454f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7464f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 747339f8db1SStefano Zampini PetscScalar vals[2]; 748339f8db1SStefano Zampini PetscInt cols[2]; 749339f8db1SStefano Zampini 750339f8db1SStefano Zampini cols[0] = idxs[i]; 7514f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 752339f8db1SStefano Zampini vals[0] = 1.; 753b0f5fe93SStefano Zampini vals[1] = 1.; 7544f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 755339f8db1SStefano Zampini } 7564f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7574f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7584f1b2e48SStefano Zampini array[nzs-1] = 1.; 7594f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7604f1b2e48SStefano Zampini /* store local idxs for p0 */ 7614f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7624f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 763339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7644f1b2e48SStefano Zampini } 765339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 766339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 767a3df083aSStefano Zampini /* project if needed */ 768a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7691dd7afcfSStefano Zampini Mat M; 7701dd7afcfSStefano Zampini 7711dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 772339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 7731dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 7741dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 775a3df083aSStefano Zampini } 7764f1b2e48SStefano Zampini /* store global idxs for p0 */ 7774f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 778339f8db1SStefano Zampini } 779ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7804f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 781b0f5fe93SStefano Zampini 782b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 783b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 784339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 785339f8db1SStefano Zampini PetscFunctionReturn(0); 786339f8db1SStefano Zampini } 787339f8db1SStefano Zampini 788339f8db1SStefano Zampini #undef __FUNCT__ 789015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 790015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 791efc2fbd9SStefano Zampini { 792efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 793efc2fbd9SStefano Zampini PetscErrorCode ierr; 794efc2fbd9SStefano Zampini 795efc2fbd9SStefano Zampini PetscFunctionBegin; 796efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 797efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 7984f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 799efc2fbd9SStefano Zampini } 800015636ebSStefano Zampini if (get) { /* use SF to get values */ 801efc2fbd9SStefano Zampini PetscScalar *array; 802efc2fbd9SStefano Zampini 803efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8044f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8054f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 806efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8074f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 8084f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 809efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 810efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 811efc2fbd9SStefano Zampini } 812efc2fbd9SStefano Zampini PetscFunctionReturn(0); 813efc2fbd9SStefano Zampini } 814efc2fbd9SStefano Zampini 815efc2fbd9SStefano Zampini #undef __FUNCT__ 816c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 817c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 818c263805aSStefano Zampini { 819c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 820c263805aSStefano Zampini PetscErrorCode ierr; 821c263805aSStefano Zampini 822c263805aSStefano Zampini PetscFunctionBegin; 823c263805aSStefano Zampini /* TODO: add error checking 824c263805aSStefano Zampini - avoid nested pop (or push) calls. 825c263805aSStefano Zampini - cannot push before pop. 8261c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 827c263805aSStefano Zampini */ 8284f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 829efc2fbd9SStefano Zampini PetscFunctionReturn(0); 830efc2fbd9SStefano Zampini } 831c263805aSStefano Zampini if (pop) { 832a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8334f1b2e48SStefano Zampini IS is_p0; 8344f1b2e48SStefano Zampini MatReuse reuse; 835c263805aSStefano Zampini 836c263805aSStefano Zampini /* extract B_0 */ 8374f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8384f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8394f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8404f1b2e48SStefano Zampini } 8414f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8424f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 843c263805aSStefano Zampini /* remove rows and cols from local problem */ 844c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 84597d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8464f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8474f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 848a3df083aSStefano Zampini } else { 849a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 850a3df083aSStefano Zampini PetscScalar *vals; 851a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 852a3df083aSStefano Zampini 853a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 854a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 855a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8560b5adadeSStefano Zampini PetscInt *nnz; 857a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 858a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 859a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 860331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 861331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 862331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 863331e053bSStefano Zampini nnz[i] = n - nnz[i]; 864331e053bSStefano Zampini } 865331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 866331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 867331e053bSStefano Zampini } 868a3df083aSStefano Zampini 869a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 870a3df083aSStefano Zampini PetscScalar *array; 871a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 872a3df083aSStefano Zampini 873a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 874a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 875a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 876a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 877a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 878a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 879a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 880a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 881a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 882a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 883a3df083aSStefano Zampini cum = 0; 884a3df083aSStefano Zampini for (j=0;j<n;j++) { 885a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 886a3df083aSStefano Zampini vals[cum] = array[j]; 887a3df083aSStefano Zampini idxs_ins[cum] = j; 888a3df083aSStefano Zampini cum++; 889a3df083aSStefano Zampini } 890a3df083aSStefano Zampini } 891a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 892a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 893a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 894a3df083aSStefano Zampini } 895a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 896a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 897a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 898a3df083aSStefano Zampini } 899c263805aSStefano Zampini } else { /* push */ 900a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9014f1b2e48SStefano Zampini PetscInt i; 9024f1b2e48SStefano Zampini 9034f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9044f1b2e48SStefano Zampini PetscScalar *B0_vals; 9054f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9064f1b2e48SStefano Zampini 9074f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9084f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9097b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9104f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9114f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9124f1b2e48SStefano Zampini } 913c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 914c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 915a3df083aSStefano Zampini } else { 916a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 917a3df083aSStefano Zampini } 918c263805aSStefano Zampini } 919c263805aSStefano Zampini PetscFunctionReturn(0); 920c263805aSStefano Zampini } 921c263805aSStefano Zampini 922c263805aSStefano Zampini #undef __FUNCT__ 923b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 92408122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 925b1b3d7a2SStefano Zampini { 926b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 92708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 92808122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 92908122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 93008122e43SStefano Zampini PetscScalar *work,lwork; 93108122e43SStefano Zampini PetscScalar *St,*S,*eigv; 93208122e43SStefano Zampini PetscScalar *Sarray,*Starray; 93308122e43SStefano Zampini PetscReal *eigs,thresh; 9341b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 935f6f667cfSStefano Zampini PetscBool allocated_S_St; 93608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 93708122e43SStefano Zampini PetscReal *rwork; 93808122e43SStefano Zampini #endif 939b1b3d7a2SStefano Zampini PetscErrorCode ierr; 940b1b3d7a2SStefano Zampini 941b1b3d7a2SStefano Zampini PetscFunctionBegin; 942af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 943af25d912SStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 94406a4e24aSStefano Zampini 945fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 946fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 947fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 948fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9491575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 950fd14bc51SStefano Zampini } 951fd14bc51SStefano Zampini 952e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 953e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 954e496cd5dSStefano Zampini } 955e496cd5dSStefano Zampini 95608122e43SStefano Zampini /* max size of subsets */ 95708122e43SStefano Zampini mss = 0; 95808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 95908122e43SStefano Zampini PetscInt subset_size; 960862806e4SStefano Zampini 96108122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 96208122e43SStefano Zampini mss = PetscMax(mss,subset_size); 96308122e43SStefano Zampini } 96408122e43SStefano Zampini 96508122e43SStefano Zampini /* min/max and threshold */ 96608122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 967f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 96808122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 969f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 970f6f667cfSStefano Zampini if (nmin) { 971f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 972f6f667cfSStefano Zampini } 97308122e43SStefano Zampini 97408122e43SStefano Zampini /* allocate lapack workspace */ 97508122e43SStefano Zampini cum = cum2 = 0; 97608122e43SStefano Zampini maxneigs = 0; 97708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 97808122e43SStefano Zampini PetscInt n,subset_size; 979f6f667cfSStefano Zampini 98008122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 98108122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9829162d606SStefano Zampini cum += subset_size; 9839162d606SStefano Zampini cum2 += subset_size*n; 98408122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 98508122e43SStefano Zampini } 98608122e43SStefano Zampini if (mss) { 9879ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 98808122e43SStefano Zampini PetscBLASInt B_itype = 1; 98908122e43SStefano Zampini PetscBLASInt B_N = mss; 9904c6709b3SStefano Zampini PetscReal zero = 0.0; 9914c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 99208122e43SStefano Zampini 99308122e43SStefano Zampini B_lwork = -1; 99408122e43SStefano Zampini S = NULL; 99508122e43SStefano Zampini St = NULL; 996a58a30b4SStefano Zampini eigs = NULL; 997a58a30b4SStefano Zampini eigv = NULL; 998a58a30b4SStefano Zampini B_iwork = NULL; 999a58a30b4SStefano Zampini B_ifail = NULL; 1000d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1001d1710679SStefano Zampini rwork = NULL; 1002d1710679SStefano Zampini #endif 10038bec7fa6SStefano Zampini thresh = 1.0; 100408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 100508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 100608122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 100708122e43SStefano Zampini #else 100808122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 100908122e43SStefano Zampini #endif 101008122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 101108122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 101208122e43SStefano Zampini } else { 101308122e43SStefano Zampini /* TODO */ 101408122e43SStefano Zampini } 101508122e43SStefano Zampini } else { 101608122e43SStefano Zampini lwork = 0; 101708122e43SStefano Zampini } 101808122e43SStefano Zampini 101908122e43SStefano Zampini nv = 0; 1020d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 1021d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 102208122e43SStefano Zampini } 10234c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1024f6f667cfSStefano Zampini if (allocated_S_St) { 1025f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1026f6f667cfSStefano Zampini } 1027f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 102808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 102908122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 103008122e43SStefano Zampini #endif 10319162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10329162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10339162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 103408122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10359162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 103608122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 103708122e43SStefano Zampini 103808122e43SStefano Zampini maxneigs = 0; 103972b8c272SStefano Zampini cum = cumarray = 0; 10409162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10419162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1042d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 104308122e43SStefano Zampini const PetscInt *idxs; 104408122e43SStefano Zampini 1045d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 104608122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 104708122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 104808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 104908122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10509162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10519162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 105208122e43SStefano Zampini } 1053d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 105408122e43SStefano Zampini } 105508122e43SStefano Zampini 105608122e43SStefano Zampini if (mss) { /* multilevel */ 105708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 105808122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 105908122e43SStefano Zampini } 106008122e43SStefano Zampini 1061ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 106208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 106308122e43SStefano Zampini const PetscInt *idxs; 10649d54b7f4SStefano Zampini PetscReal upper,lower; 1065862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 106608122e43SStefano Zampini PetscBLASInt B_N; 1067aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 106808122e43SStefano Zampini 10699d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 10709d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 10719d54b7f4SStefano Zampini lower = thresh; 10729d54b7f4SStefano Zampini } else { 10739d54b7f4SStefano Zampini upper = 1./thresh; 10749d54b7f4SStefano Zampini lower = 0.; 10759d54b7f4SStefano Zampini } 1076862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1077ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1078f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1079f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1081aff50787SStefano Zampini PetscInt j,k; 1082aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1083aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1084aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 108508122e43SStefano Zampini } 108608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1087aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1088aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1089aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1090aff50787SStefano Zampini } 109108122e43SStefano Zampini } 109208122e43SStefano Zampini } else { 109308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109508122e43SStefano Zampini } 10968bec7fa6SStefano Zampini } else { 1097f6f667cfSStefano Zampini S = Sarray + cumarray; 1098f6f667cfSStefano Zampini St = Starray + cumarray; 10998bec7fa6SStefano Zampini } 1100aff50787SStefano Zampini /* see if we can save some work */ 1101*b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1102aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1103aff50787SStefano Zampini } 1104aff50787SStefano Zampini 1105*b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1106aff50787SStefano Zampini B_neigs = 0; 1107aff50787SStefano Zampini } else { 11089ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 110908122e43SStefano Zampini PetscBLASInt B_itype = 1; 1110f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11114c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11129552c7c7SStefano Zampini PetscInt nmin_s; 1113*b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 111408122e43SStefano Zampini 1115fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11168bec7fa6SStefano 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]]); 1117fd14bc51SStefano Zampini } 1118d16cbb6bSStefano Zampini 1119*b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1120*b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1121*b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1122*b7ab4a40SStefano Zampini } 1123*b7ab4a40SStefano Zampini 112408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1125*b7ab4a40SStefano Zampini if (compute_range) { 1126d16cbb6bSStefano Zampini 1127d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 112808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11299d54b7f4SStefano 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)); 113008122e43SStefano Zampini #else 11319d54b7f4SStefano 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)); 113208122e43SStefano Zampini #endif 1133*b7ab4a40SStefano Zampini } else if (!same_data) { 1134d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1135d16cbb6bSStefano Zampini B_IL = 1; 1136d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 11379d54b7f4SStefano 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)); 1138d16cbb6bSStefano Zampini #else 11399d54b7f4SStefano 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)); 1140d16cbb6bSStefano Zampini #endif 1141*b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1142*b7ab4a40SStefano Zampini PetscInt k; 1143*b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1144*b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1145*b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1146*b7ab4a40SStefano Zampini nmin = nmax; 1147*b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1148*b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1149*b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1150*b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1151*b7ab4a40SStefano Zampini } 1152d16cbb6bSStefano Zampini } 115308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 115408122e43SStefano Zampini if (B_ierr) { 11556c4ed002SBarry 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); 11566c4ed002SBarry 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); 11576c4ed002SBarry 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); 115808122e43SStefano Zampini } 115908122e43SStefano Zampini 116008122e43SStefano Zampini if (B_neigs > nmax) { 1161fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1162fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1163fd14bc51SStefano Zampini } 11649d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 116508122e43SStefano Zampini B_neigs = nmax; 116608122e43SStefano Zampini } 116708122e43SStefano Zampini 11689552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11699552c7c7SStefano Zampini if (B_neigs < nmin_s) { 117008122e43SStefano Zampini PetscBLASInt B_neigs2; 117108122e43SStefano Zampini 11729d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1173f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 11749d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 11759d54b7f4SStefano Zampini } else { 11769d54b7f4SStefano Zampini B_IL = B_neigs + 1; 11779d54b7f4SStefano Zampini B_IU = nmin_s; 11789d54b7f4SStefano Zampini } 1179fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1180fd14bc51SStefano 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); 1181fd14bc51SStefano Zampini } 11829ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 118308122e43SStefano Zampini PetscInt j; 118408122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118508122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118608122e43SStefano Zampini } 118708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118808122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118908122e43SStefano Zampini } 119008122e43SStefano Zampini } else { 119108122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119208122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119308122e43SStefano Zampini } 119408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 119508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11969d54b7f4SStefano 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)); 119708122e43SStefano Zampini #else 11989d54b7f4SStefano 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)); 119908122e43SStefano Zampini #endif 120008122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 120108122e43SStefano Zampini B_neigs += B_neigs2; 120208122e43SStefano Zampini } 120308122e43SStefano Zampini if (B_ierr) { 12046c4ed002SBarry 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); 12056c4ed002SBarry 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); 12066c4ed002SBarry 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); 120708122e43SStefano Zampini } 1208fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1209ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 121008122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 121108122e43SStefano Zampini if (eigs[j] == 0.0) { 1212ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 121308122e43SStefano Zampini } else { 12149d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1215ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 12169d54b7f4SStefano Zampini } else { 12179d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 12189d54b7f4SStefano Zampini } 1219fd14bc51SStefano Zampini } 122008122e43SStefano Zampini } 122108122e43SStefano Zampini } 122208122e43SStefano Zampini } else { 122308122e43SStefano Zampini /* TODO */ 122408122e43SStefano Zampini } 1225aff50787SStefano Zampini } 12266c3e6151SStefano Zampini /* change the basis back to the original one */ 12276c3e6151SStefano Zampini if (sub_schurs->change) { 122872b8c272SStefano Zampini Mat change,phi,phit; 12296c3e6151SStefano Zampini 12306c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12316c3e6151SStefano Zampini PetscInt ii; 12326c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12336c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12346c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12356c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12366c3e6151SStefano Zampini } 12376c3e6151SStefano Zampini } 12386c3e6151SStefano Zampini } 123972b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12406c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 124172b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12426c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12436c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12446c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12456c3e6151SStefano Zampini } 12468bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12478bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12489162d606SStefano Zampini if (B_neigs) { 12499162d606SStefano 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); 1250fd14bc51SStefano Zampini 1251fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12529552c7c7SStefano Zampini PetscInt ii; 12539552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1254ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12559552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1256ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1257ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1258ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1259ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1260ac47001eSStefano Zampini #else 1261ac47001eSStefano 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); 1262ac47001eSStefano Zampini #endif 12639552c7c7SStefano Zampini } 12649552c7c7SStefano Zampini } 1265fd14bc51SStefano Zampini } 12669162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12679162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12689162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12699162d606SStefano Zampini cum++; 127008122e43SStefano Zampini } 127108122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 127208122e43SStefano Zampini /* shift for next computation */ 127308122e43SStefano Zampini cumarray += subset_size*subset_size; 127408122e43SStefano Zampini } 1275fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1276fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1277fd14bc51SStefano Zampini } 127808122e43SStefano Zampini 127908122e43SStefano Zampini if (mss) { 128008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 128108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1282f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1283f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1284f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 128508122e43SStefano Zampini } 1286f6f667cfSStefano Zampini if (allocated_S_St) { 1287f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1288f6f667cfSStefano Zampini } 1289f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 129008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 129108122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 129208122e43SStefano Zampini #endif 129308122e43SStefano Zampini if (pcbddc->dbg_flag) { 12941b968477SStefano Zampini PetscInt maxneigs_r; 1295b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12969b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 129708122e43SStefano Zampini } 129808122e43SStefano Zampini PetscFunctionReturn(0); 129908122e43SStefano Zampini } 1300b1b3d7a2SStefano Zampini 1301674ae819SStefano Zampini #undef __FUNCT__ 1302c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1303c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1304c8587f34SStefano Zampini { 1305c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13068629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1307c8587f34SStefano Zampini PetscErrorCode ierr; 1308c8587f34SStefano Zampini 1309c8587f34SStefano Zampini PetscFunctionBegin; 1310f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 13115e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1312c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1313c8587f34SStefano Zampini 1314684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 13150fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1316684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1317c8587f34SStefano Zampini 1318c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1319b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1320c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1321c8587f34SStefano Zampini } 1322c8587f34SStefano Zampini 13238629588bSStefano Zampini /* 13248629588bSStefano Zampini Setup local correction and local part of coarse basis. 13258629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13268629588bSStefano Zampini */ 132747f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13288629588bSStefano Zampini 13298629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13308629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13318629588bSStefano Zampini 13328629588bSStefano Zampini /* free */ 13338629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1334c8587f34SStefano Zampini PetscFunctionReturn(0); 1335c8587f34SStefano Zampini } 1336c8587f34SStefano Zampini 1337c8587f34SStefano Zampini #undef __FUNCT__ 1338674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1339674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1340674ae819SStefano Zampini { 1341674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1342674ae819SStefano Zampini PetscErrorCode ierr; 1343674ae819SStefano Zampini 1344674ae819SStefano Zampini PetscFunctionBegin; 1345674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1346674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 134730368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1348674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1349674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1350785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1351674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1352f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1353f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1354785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 135563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 135663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1357674ae819SStefano Zampini PetscFunctionReturn(0); 1358674ae819SStefano Zampini } 1359674ae819SStefano Zampini 1360674ae819SStefano Zampini #undef __FUNCT__ 1361674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1362674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1363674ae819SStefano Zampini { 1364674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13654f1b2e48SStefano Zampini PetscInt i; 1366674ae819SStefano Zampini PetscErrorCode ierr; 1367674ae819SStefano Zampini 1368674ae819SStefano Zampini PetscFunctionBegin; 1369b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1370674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 137116909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 13721dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1373674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1374674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13754f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13764f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13774f1b2e48SStefano Zampini } 13784f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1379b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1380674ae819SStefano Zampini PetscFunctionReturn(0); 1381674ae819SStefano Zampini } 1382674ae819SStefano Zampini 1383674ae819SStefano Zampini #undef __FUNCT__ 1384674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1385674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1386674ae819SStefano Zampini { 1387674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1388674ae819SStefano Zampini PetscErrorCode ierr; 1389674ae819SStefano Zampini 1390674ae819SStefano Zampini PetscFunctionBegin; 1391674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 139258da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1393ca92afb2SStefano Zampini PetscScalar *array; 139406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 139506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 139658da7f69SStefano Zampini } 1397674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1398674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 139915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 140015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1401674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1402674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1403674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 140406656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1405674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1406674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14078ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1408674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1409674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1410674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1411f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1412f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1413f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1414f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1415727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 14160e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1417f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 141870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 14196e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 142081d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14210369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14221dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 14234f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14248b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1425ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1426ca92afb2SStefano Zampini PetscInt i; 1427ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1428ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1429ca92afb2SStefano Zampini } 1430ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1431ca92afb2SStefano Zampini } 14324f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1433674ae819SStefano Zampini PetscFunctionReturn(0); 1434674ae819SStefano Zampini } 1435674ae819SStefano Zampini 1436674ae819SStefano Zampini #undef __FUNCT__ 1437f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1438f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14396bfb1811SStefano Zampini { 14406bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14416bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14426bfb1811SStefano Zampini VecType impVecType; 14434f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14446bfb1811SStefano Zampini PetscErrorCode ierr; 14456bfb1811SStefano Zampini 14466bfb1811SStefano Zampini PetscFunctionBegin; 14476c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1448e7b262bdSStefano Zampini /* get sizes */ 14494f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1450b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14516bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1452e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1453e7b262bdSStefano Zampini /* R nodes */ 1454e7b262bdSStefano Zampini old_size = -1; 1455e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1456e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1457e7b262bdSStefano Zampini } 1458e7b262bdSStefano Zampini if (n_R != old_size) { 1459e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1460e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14616bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14626bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14636bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14646bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1465e7b262bdSStefano Zampini } 1466e7b262bdSStefano Zampini /* local primal dofs */ 1467e7b262bdSStefano Zampini old_size = -1; 1468e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1469e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1470e7b262bdSStefano Zampini } 1471e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1472e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 147383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1474e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14756bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1476e7b262bdSStefano Zampini } 1477e7b262bdSStefano Zampini /* local explicit constraints */ 1478e7b262bdSStefano Zampini old_size = -1; 1479e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1480e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1481e7b262bdSStefano Zampini } 1482e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1483e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 148483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 148583b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 148683b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 148783b7ccabSStefano Zampini } 14886bfb1811SStefano Zampini PetscFunctionReturn(0); 14896bfb1811SStefano Zampini } 14906bfb1811SStefano Zampini 14916bfb1811SStefano Zampini #undef __FUNCT__ 149247f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 149347f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 149488ebb749SStefano Zampini { 149525084f0cSStefano Zampini PetscErrorCode ierr; 149625084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 149788ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 149888ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1499d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 150025084f0cSStefano Zampini /* submatrices of local problem */ 150180677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 150206656605SStefano Zampini /* submatrices of local coarse problem */ 150306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 150425084f0cSStefano Zampini /* working matrices */ 150506656605SStefano Zampini Mat C_CR; 150625084f0cSStefano Zampini /* additional working stuff */ 150706656605SStefano Zampini PC pc_R; 15084f1b2e48SStefano Zampini Mat F; 15095cbda25cSStefano Zampini Vec dummy_vec; 1510a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 151125084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 151206656605SStefano Zampini PetscScalar *work; 151306656605SStefano Zampini PetscInt *idx_V_B; 1514ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 151506656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1516ffd830a3SStefano Zampini 151725084f0cSStefano Zampini /* some shortcuts to scalars */ 151806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 151988ebb749SStefano Zampini 152088ebb749SStefano Zampini PetscFunctionBegin; 1521ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1522ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1523ffd830a3SStefano Zampini } 1524ffd830a3SStefano Zampini 1525ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1526b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15274f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1528b371cd4fSStefano Zampini n_B = pcis->n_B; 1529b371cd4fSStefano Zampini n_D = pcis->n - n_B; 153088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 153188ebb749SStefano Zampini 153288ebb749SStefano Zampini /* vertices in boundary numbering */ 1533785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15340e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 15356c4ed002SBarry 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); 153688ebb749SStefano Zampini 153706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1538019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 153906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 154006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 154106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 154206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 154306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 154406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 154506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 154606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 154706656605SStefano Zampini 154806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 154906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 155106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 155206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1553ffd830a3SStefano Zampini lda_rhs = n_R; 1554a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 155506656605SStefano Zampini if (isLU || isILU || isCHOL) { 155606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1557df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1558df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1559d62866d3SStefano Zampini MatFactorType type; 1560d62866d3SStefano Zampini 1561df4d28bfSStefano Zampini F = reuse_solver->F; 15626816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1563d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1564ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1565df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 156606656605SStefano Zampini } else { 156706656605SStefano Zampini F = NULL; 156806656605SStefano Zampini } 156906656605SStefano Zampini 1570ffd830a3SStefano Zampini /* allocate workspace */ 1571ffd830a3SStefano Zampini n = 0; 1572ffd830a3SStefano Zampini if (n_constraints) { 1573ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1574ffd830a3SStefano Zampini } 1575ffd830a3SStefano Zampini if (n_vertices) { 1576ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1577ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1578ffd830a3SStefano Zampini } 1579ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1580ffd830a3SStefano Zampini 15815cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 15825cbda25cSStefano Zampini dummy_vec = NULL; 15835cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 15845cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 15855cbda25cSStefano Zampini } 15865cbda25cSStefano Zampini 158788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 158888ebb749SStefano Zampini if (n_constraints) { 158972b8c272SStefano Zampini Mat M1,M2,M3,C_B; 159006656605SStefano Zampini IS is_aux; 159180677318SStefano Zampini PetscScalar *array,*array2; 159206656605SStefano Zampini 1593f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 159480677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 159588ebb749SStefano Zampini 159625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 159725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15988ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 159972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 160088ebb749SStefano Zampini 160180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 160280677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1603ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 160488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 160506656605SStefano Zampini const PetscScalar *row_cmat_values; 160606656605SStefano Zampini const PetscInt *row_cmat_indices; 160706656605SStefano Zampini PetscInt size_of_constraint,j; 160888ebb749SStefano Zampini 160906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1611ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 161206656605SStefano Zampini } 161306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161406656605SStefano Zampini } 1615ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 161606656605SStefano Zampini if (F) { 161706656605SStefano Zampini Mat B; 161806656605SStefano Zampini 1619ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1620a3df083aSStefano Zampini if (need_benign_correction) { 1621df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1622a3df083aSStefano Zampini 162372b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 162472b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1625a3df083aSStefano Zampini } 162680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1627a3df083aSStefano Zampini if (need_benign_correction) { 1628a3df083aSStefano Zampini PetscScalar *marr; 1629df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1630a3df083aSStefano Zampini 1631a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 16325cbda25cSStefano Zampini if (lda_rhs != n_R) { 16335cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 16345cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 16355cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 16365cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 16375cbda25cSStefano Zampini } 16385cbda25cSStefano Zampini } else { 1639a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1640a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 16415cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1642a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1643a3df083aSStefano Zampini } 16445cbda25cSStefano Zampini } 1645a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1646a3df083aSStefano Zampini } 164706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 164806656605SStefano Zampini } else { 164980677318SStefano Zampini PetscScalar *marr; 165080677318SStefano Zampini 165180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 165206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1653ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1654ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 165506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 165606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 165706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 165806656605SStefano Zampini } 165980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 166006656605SStefano Zampini } 166180677318SStefano Zampini if (!pcbddc->switch_static) { 166280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 166380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 166480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 166580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1666ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 166780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 166880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 167180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 167280677318SStefano Zampini } 167380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 167480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 167572b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 167680677318SStefano Zampini } else { 1677ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1678ffd830a3SStefano Zampini IS dummy; 1679ffd830a3SStefano Zampini 1680ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 168172b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1682ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1683ffd830a3SStefano Zampini } else { 168480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 168580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1686ffd830a3SStefano Zampini } 168725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 168880677318SStefano Zampini } 168980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 169080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 169180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 169206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 169306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 169480677318SStefano Zampini if (isCHOL) { 169580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 169680677318SStefano Zampini } else { 169725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 169880677318SStefano Zampini } 169980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 170006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 170125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 170225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 170325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 170480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 170572b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 170672b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 170706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 170806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1709f4ddd8eeSStefano Zampini } 1710fc227af8SStefano Zampini 1711fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 171288ebb749SStefano Zampini if (n_vertices) { 171306656605SStefano Zampini IS is_aux; 17143a50541eSStefano Zampini 1715df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 17166816873aSStefano Zampini IS tis; 17176816873aSStefano Zampini 17186816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17196816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17206816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17216816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17226816873aSStefano Zampini } else { 17233a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17246816873aSStefano Zampini } 17259577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17269577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 172704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 172825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 172988ebb749SStefano Zampini } 173088ebb749SStefano Zampini 173188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1732f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 173306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 173406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 173506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 173606656605SStefano Zampini } 1737f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 173806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 173906656605SStefano Zampini PetscScalar *marray; 174006656605SStefano Zampini 174106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 174206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1743f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1744f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1745f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1746f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1747f4ddd8eeSStefano Zampini } 1748f4ddd8eeSStefano Zampini } 174906656605SStefano Zampini 1750f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 175106656605SStefano Zampini PetscScalar *marray; 175288ebb749SStefano Zampini 175306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17548eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 175506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 175688ebb749SStefano Zampini } 17573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 175806656605SStefano Zampini n *= 2; 175988ebb749SStefano Zampini } 176006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 176106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 176206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17638eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 176406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 176506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 176688ebb749SStefano Zampini } 17673301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 176806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17698eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 177006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 177106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 177288ebb749SStefano Zampini } 177388ebb749SStefano Zampini } else { 1774c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1775c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17761b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1777c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1778c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1779c0553b1fSStefano Zampini } 178088ebb749SStefano Zampini } 178106656605SStefano Zampini } 1782019a44ceSStefano Zampini 178306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17844f1b2e48SStefano Zampini p0_lidx_I = NULL; 17854f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1786d12edf2fSStefano Zampini const PetscInt *idxs; 1787d12edf2fSStefano Zampini 1788d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17894f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17904f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17914f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17924f1b2e48SStefano Zampini } 1793d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1794d12edf2fSStefano Zampini } 1795d16cbb6bSStefano Zampini 179606656605SStefano Zampini /* vertices */ 179706656605SStefano Zampini if (n_vertices) { 179816f15bc4SStefano Zampini 1799af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 180004708bb6SStefano Zampini 180116f15bc4SStefano Zampini if (n_R) { 180214393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 180306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 180416f15bc4SStefano Zampini PetscScalar *x,*y; 180504708bb6SStefano Zampini PetscBool isseqaij; 180606656605SStefano Zampini 180721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 180814393ed6SStefano Zampini if (need_benign_correction) { 180914393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 181014393ed6SStefano Zampini IS is_p0; 181114393ed6SStefano Zampini PetscInt *idxs_p0,n; 181214393ed6SStefano Zampini 181314393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 181414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 181514393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 1816af25d912SStefano 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); 181714393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 181814393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 181914393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 182014393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 182114393ed6SStefano Zampini } 182214393ed6SStefano Zampini 1823ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1824af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1825ffd830a3SStefano Zampini } else { 1826ca92afb2SStefano Zampini PetscScalar *av,*array; 1827ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1828ca92afb2SStefano Zampini PetscInt n; 1829ca92afb2SStefano Zampini PetscBool flg_row; 1830ffd830a3SStefano Zampini 1831ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1832ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 18339d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1834ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1835ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1836ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1837ca92afb2SStefano Zampini PetscInt j; 1838ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1839ffd830a3SStefano Zampini } 1840ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1841ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1842ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1843ffd830a3SStefano Zampini } 1844ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1845a3df083aSStefano Zampini if (need_benign_correction) { 1846df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1847a3df083aSStefano Zampini PetscScalar *marr; 1848a3df083aSStefano Zampini 1849a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 185014393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 185114393ed6SStefano Zampini 185214393ed6SStefano Zampini | 0 0 0 | (V) 185314393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 185414393ed6SStefano Zampini | 0 0 -1 | (p0) 185514393ed6SStefano Zampini 185614393ed6SStefano Zampini */ 1857df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 185814393ed6SStefano Zampini const PetscScalar *vals; 185914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 186014393ed6SStefano Zampini PetscInt n,j,nz; 186114393ed6SStefano Zampini 1862df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1863df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186414393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 186514393ed6SStefano Zampini for (j=0;j<n;j++) { 186614393ed6SStefano Zampini PetscScalar val = vals[j]; 186714393ed6SStefano Zampini PetscInt k,col = idxs[j]; 186814393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 186914393ed6SStefano Zampini } 187014393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1871df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 187214393ed6SStefano Zampini } 187372b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 187472b8c272SStefano Zampini } 187572b8c272SStefano Zampini if (F) { 187614393ed6SStefano Zampini /* need to correct the rhs */ 187772b8c272SStefano Zampini if (need_benign_correction) { 187872b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 187972b8c272SStefano Zampini PetscScalar *marr; 188072b8c272SStefano Zampini 188172b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 18825cbda25cSStefano Zampini if (lda_rhs != n_R) { 18835cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18845cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18855cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 18865cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18875cbda25cSStefano Zampini } 18885cbda25cSStefano Zampini } else { 1889a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1890a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18915cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1892a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1893a3df083aSStefano Zampini } 18945cbda25cSStefano Zampini } 1895a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1896a3df083aSStefano Zampini } 189706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 189814393ed6SStefano Zampini /* need to correct the solution */ 1899a3df083aSStefano Zampini if (need_benign_correction) { 1900df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1901a3df083aSStefano Zampini PetscScalar *marr; 1902a3df083aSStefano Zampini 1903a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 19045cbda25cSStefano Zampini if (lda_rhs != n_R) { 19055cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 19065cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19075cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19085cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19095cbda25cSStefano Zampini } 19105cbda25cSStefano Zampini } else { 1911a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1912a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19135cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1914a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1915a3df083aSStefano Zampini } 19165cbda25cSStefano Zampini } 1917a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1918a3df083aSStefano Zampini } 191906656605SStefano Zampini } else { 192006656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 192106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1922ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1923ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 192406656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 192506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 192606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 192706656605SStefano Zampini } 192806656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 192906656605SStefano Zampini } 193080677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1931ffd830a3SStefano Zampini /* S_VV and S_CV */ 193206656605SStefano Zampini if (n_constraints) { 193306656605SStefano Zampini Mat B; 193480677318SStefano Zampini 1935ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 193680677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1937ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1938ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 193980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194180677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 194280677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 194380677318SStefano Zampini } 1944ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 194580677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 194680677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1947ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 194880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 194906656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1950ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1951ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 195206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195306656605SStefano Zampini } 195404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 195504708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 1956511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 195704708bb6SStefano Zampini } 1958ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1959ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1960ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1961ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1962ffd830a3SStefano Zampini } 196306656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 196414393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 196514393ed6SStefano Zampini if (need_benign_correction) { 1966df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 196714393ed6SStefano Zampini PetscScalar *marr,*sums; 196814393ed6SStefano Zampini 196914393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 197014393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1971df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 197214393ed6SStefano Zampini const PetscScalar *vals; 197314393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 197414393ed6SStefano Zampini PetscInt n,j,nz; 197514393ed6SStefano Zampini 1976df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1977df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 197814393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 197914393ed6SStefano Zampini PetscInt k; 198014393ed6SStefano Zampini sums[j] = 0.; 198114393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 198214393ed6SStefano Zampini } 198314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 198414393ed6SStefano Zampini for (j=0;j<n;j++) { 198514393ed6SStefano Zampini PetscScalar val = vals[j]; 198614393ed6SStefano Zampini PetscInt k; 198714393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 198814393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 198914393ed6SStefano Zampini } 199014393ed6SStefano Zampini } 199114393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1992df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 199314393ed6SStefano Zampini } 199414393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 199514393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 199614393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 199714393ed6SStefano Zampini } 199880677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 199906656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 200006656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 200206656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 200306656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 200406656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 200506656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2006d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2007019a44ceSStefano Zampini } else { 2008d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2009d16cbb6bSStefano Zampini } 201021eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2011d16cbb6bSStefano Zampini 201206656605SStefano Zampini /* coarse basis functions */ 201306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 201416f15bc4SStefano Zampini PetscScalar *y; 201516f15bc4SStefano Zampini 2016ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 201706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 201806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 201906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 202206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 202406656605SStefano Zampini 202506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20264f1b2e48SStefano Zampini PetscInt j; 20274f1b2e48SStefano Zampini 202806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 202906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 203006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20334f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 203406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203506656605SStefano Zampini } 203606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 203706656605SStefano Zampini } 203804708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 203904708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 204006656605SStefano Zampini } 20415cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 204206656605SStefano Zampini 204306656605SStefano Zampini if (n_constraints) { 204406656605SStefano Zampini Mat B; 204506656605SStefano Zampini 2046ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 204706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 204880677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 204906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205006656605SStefano Zampini if (n_vertices) { 205180677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 205280677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 205380677318SStefano Zampini } else { 205480677318SStefano Zampini Mat S_VCt; 205580677318SStefano Zampini 2056ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2057ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 205872b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2059ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2060ffd830a3SStefano Zampini } 206180677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 206280677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 206380677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 206480677318SStefano Zampini } 206506656605SStefano Zampini } 206606656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206706656605SStefano Zampini /* coarse basis functions */ 206806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 206906656605SStefano Zampini PetscScalar *y; 207006656605SStefano Zampini 2071ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 207206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 207406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 207806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20794f1b2e48SStefano Zampini PetscInt j; 20804f1b2e48SStefano Zampini 208106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 208206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 208306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20864f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 208706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 208806656605SStefano Zampini } 208906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209006656605SStefano Zampini } 209106656605SStefano Zampini } 209280677318SStefano Zampini if (n_constraints) { 209380677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 209480677318SStefano Zampini } 20954f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 209672b8c272SStefano Zampini 209772b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 209872b8c272SStefano Zampini if (pcbddc->benign_n) { 209972b8c272SStefano Zampini Mat B0_B,B0_BPHI; 210072b8c272SStefano Zampini IS is_dummy; 210172b8c272SStefano Zampini PetscScalar *data; 210272b8c272SStefano Zampini PetscInt j; 210372b8c272SStefano Zampini 210472b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 210572b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 210672b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 210772b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 210886c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 210972b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 211072b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 211172b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 211272b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 211372b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 211472b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 211572b8c272SStefano Zampini } 211672b8c272SStefano Zampini } 211772b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 211872b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 211972b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 212072b8c272SStefano Zampini } 2121019a44ceSStefano Zampini 212206656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 21233301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2124ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2125ffd830a3SStefano Zampini PetscScalar *marray; 212606656605SStefano Zampini 212706656605SStefano Zampini if (n_constraints) { 2128ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 212906656605SStefano Zampini 2130af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 213106656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2132ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 213316f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 213406656605SStefano Zampini if (n_vertices) { 2135ffd830a3SStefano Zampini Mat S_VCT; 213606656605SStefano Zampini 213706656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2138ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 213916f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 214006656605SStefano Zampini } 2141ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 214206656605SStefano Zampini } 214316f15bc4SStefano Zampini if (n_vertices && n_R) { 2144ffd830a3SStefano Zampini PetscScalar *av,*marray; 2145ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2146ffd830a3SStefano Zampini PetscInt n; 2147ffd830a3SStefano Zampini PetscBool flg_row; 214806656605SStefano Zampini 2149ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2150af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2151ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2152ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2153ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2154ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2155ffd830a3SStefano Zampini PetscInt j; 2156ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2157ffd830a3SStefano Zampini } 2158ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2159ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2160ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 216106656605SStefano Zampini } 216206656605SStefano Zampini 2163ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2164ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2165ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2166ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2167ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 216806656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 216906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 217106656605SStefano Zampini } 2172ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2173ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2174ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2175ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2176ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2177ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2178ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2179ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 218006656605SStefano Zampini } 2181ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 218206656605SStefano Zampini /* coarse basis functions */ 218306656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 218406656605SStefano Zampini PetscScalar *y; 218506656605SStefano Zampini 2186ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 218706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 218806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 218906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219106656605SStefano Zampini if (i<n_vertices) { 219206656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 219306656605SStefano Zampini } 219406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 219606656605SStefano Zampini 219706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 219806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 219906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 220006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 220306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220406656605SStefano Zampini } 220506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 220606656605SStefano Zampini } 2207ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2208ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 220906656605SStefano Zampini } 2210d62866d3SStefano Zampini /* free memory */ 221188ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 221206656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 221306656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 221406656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 221506656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2216d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2217d62866d3SStefano Zampini if (n_vertices) { 2218d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2219d62866d3SStefano Zampini } 2220d62866d3SStefano Zampini if (n_constraints) { 2221d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2222d62866d3SStefano Zampini } 222388ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 222488ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 222588ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2226d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 222788ebb749SStefano Zampini Mat coarse_sub_mat; 222825084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 222988ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 223088ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 223188ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22328bec7fa6SStefano Zampini Mat C_B,CPHI; 22338bec7fa6SStefano Zampini IS is_dummy; 22348bec7fa6SStefano Zampini Vec mones; 223588ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 223688ebb749SStefano Zampini PetscReal real_value; 223788ebb749SStefano Zampini 2238a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2239a3df083aSStefano Zampini Mat A; 2240a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2241a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2242a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2243a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2244a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2245a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2246a3df083aSStefano Zampini } else { 224788ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 224888ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 224988ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 225088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2251a3df083aSStefano Zampini } 225288ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 225388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2254ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 225588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 225688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 225788ebb749SStefano Zampini } 225888ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 225988ebb749SStefano Zampini 226025084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22613301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 226225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2263ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226488ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 226588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 226688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 226788ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 226888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 226988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 227288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227688ebb749SStefano Zampini } else { 227788ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 227888ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 227988ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 228188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228288ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 228488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228588ebb749SStefano Zampini } 228688ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 228788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 228888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2289511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 22904f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2291fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2292d12edf2fSStefano Zampini PetscScalar *data,*data2; 22934f1b2e48SStefano Zampini PetscInt j; 2294d12edf2fSStefano Zampini 22954f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2296fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2297d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 229886c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2299d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2300d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 23014f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 23024f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2303d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 23044f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 23054f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 23064f1b2e48SStefano Zampini } 2307d12edf2fSStefano Zampini } 2308d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2309d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2310d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2311d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2312d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2313d12edf2fSStefano Zampini } 2314d12edf2fSStefano Zampini #if 0 2315d12edf2fSStefano Zampini { 2316d12edf2fSStefano Zampini PetscViewer viewer; 2317d12edf2fSStefano Zampini char filename[256]; 2318ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2319d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2320d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2321ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2322ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2323ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2324d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 232572b8c272SStefano Zampini if (save_change) { 232672b8c272SStefano Zampini Mat phi_B; 232772b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 232872b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 232972b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 233072b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 233172b8c272SStefano Zampini } else { 2332ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2333ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 233472b8c272SStefano Zampini } 2335ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2336ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2337ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2338ffd830a3SStefano Zampini } 2339ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2340ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2341ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2342ffd830a3SStefano Zampini } 234372b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2344ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2345ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2346ffd830a3SStefano Zampini } 2347d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2348d12edf2fSStefano Zampini } 2349d12edf2fSStefano Zampini #endif 235081d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23518bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23521575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 235306656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23548bec7fa6SStefano Zampini 23558bec7fa6SStefano Zampini /* check constraints */ 2356a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2357a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 23584f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23598bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2360a00504b5SStefano Zampini } else { 2361a00504b5SStefano Zampini PetscScalar *data; 2362a00504b5SStefano Zampini Mat tmat; 2363a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2364a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2365a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2366a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2367a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2368a00504b5SStefano Zampini } 23698bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23708bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23718bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23728bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2373bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2374ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2375bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2376bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2377bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2378bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2379bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 238088ebb749SStefano Zampini } 23818bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23828bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23838bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23848bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 238525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 238688ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 238788ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 238888ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 238988ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 239088ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 239188ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 239288ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 239388ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 239488ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 239588ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2396ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 239788ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 239888ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 239988ebb749SStefano Zampini } 240088ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 240188ebb749SStefano Zampini } 24028629588bSStefano Zampini /* get back data */ 24038629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 240488ebb749SStefano Zampini PetscFunctionReturn(0); 240588ebb749SStefano Zampini } 240688ebb749SStefano Zampini 240788ebb749SStefano Zampini #undef __FUNCT__ 2408d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2409d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2410aa0d41d4SStefano Zampini { 2411d65f70fdSStefano Zampini Mat *work_mat; 2412d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2413d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2414d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2415aa0d41d4SStefano Zampini PetscErrorCode ierr; 2416aa0d41d4SStefano Zampini 2417aa0d41d4SStefano Zampini PetscFunctionBegin; 2418d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2419d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2420d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2421d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2422aa0d41d4SStefano Zampini 2423d65f70fdSStefano Zampini if (!rsorted) { 2424906d46d4SStefano Zampini const PetscInt *idxs; 2425906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2426aa0d41d4SStefano Zampini 2427d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2428d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2429d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2430d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2431aa0d41d4SStefano Zampini } 2432d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2433d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2434d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2435d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2436aa0d41d4SStefano Zampini } 2437d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2438d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2439d65f70fdSStefano Zampini } else { 2440d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2441d65f70fdSStefano Zampini isrow_s = isrow; 2442aa0d41d4SStefano Zampini } 2443906d46d4SStefano Zampini 2444d65f70fdSStefano Zampini if (!csorted) { 2445d65f70fdSStefano Zampini if (isrow == iscol) { 2446d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2447d65f70fdSStefano Zampini iscol_s = isrow_s; 2448d65f70fdSStefano Zampini } else { 2449d65f70fdSStefano Zampini const PetscInt *idxs; 2450d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2451906d46d4SStefano Zampini 2452d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2453d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2454d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2455d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2456d65f70fdSStefano Zampini } 2457d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2458d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2459d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2460d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2461d65f70fdSStefano Zampini } 2462d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2463d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2464d65f70fdSStefano Zampini } 2465d65f70fdSStefano Zampini } else { 2466d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2467d65f70fdSStefano Zampini iscol_s = iscol; 2468d65f70fdSStefano Zampini } 2469d65f70fdSStefano Zampini 2470d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2471d65f70fdSStefano Zampini 2472d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2473906d46d4SStefano Zampini Mat new_mat; 2474d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2475906d46d4SStefano Zampini 2476d65f70fdSStefano Zampini if (!rsorted) { 2477d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2478d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2479d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2480d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2481906d46d4SStefano Zampini } 2482d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2483d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2484d65f70fdSStefano Zampini } else { 2485d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2486906d46d4SStefano Zampini } 2487d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2488d65f70fdSStefano Zampini 2489d65f70fdSStefano Zampini if (!csorted) { 2490d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2491d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2492d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2493d65f70fdSStefano Zampini } else { 2494d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2495d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2496d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2497d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2498d65f70fdSStefano Zampini } 2499d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2500d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2501d65f70fdSStefano Zampini } 2502d65f70fdSStefano Zampini } else { 2503d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2504d65f70fdSStefano Zampini } 2505d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2506d65f70fdSStefano Zampini 2507d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2508d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2509d65f70fdSStefano Zampini work_mat[0] = new_mat; 2510d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2511d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2512d65f70fdSStefano Zampini } 2513d65f70fdSStefano Zampini 2514d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2515d65f70fdSStefano Zampini *B = work_mat[0]; 2516d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2517d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2518d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2519d65f70fdSStefano Zampini PetscFunctionReturn(0); 2520d65f70fdSStefano Zampini } 2521d65f70fdSStefano Zampini 2522d65f70fdSStefano Zampini #undef __FUNCT__ 25235e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 25245e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2525aa0d41d4SStefano Zampini { 2526aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 25275e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2528d65f70fdSStefano Zampini Mat new_mat; 25295e8657edSStefano Zampini IS is_local,is_global; 2530d65f70fdSStefano Zampini PetscInt local_size; 2531d65f70fdSStefano Zampini PetscBool isseqaij; 2532aa0d41d4SStefano Zampini PetscErrorCode ierr; 2533aa0d41d4SStefano Zampini 2534aa0d41d4SStefano Zampini PetscFunctionBegin; 2535aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25365e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25375e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2538b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2539aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2540d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2541aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2542906d46d4SStefano Zampini 2543906d46d4SStefano Zampini /* check */ 2544906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2545906d46d4SStefano Zampini Vec x,x_change; 2546906d46d4SStefano Zampini PetscReal error; 2547906d46d4SStefano Zampini 25485e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2549906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25505e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2551e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2552e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2553d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2554e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2555e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2556906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2557906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2558906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2559906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2560906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2561906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2562906d46d4SStefano Zampini } 2563906d46d4SStefano Zampini 256422d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25659b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 256622d5777bSStefano Zampini if (isseqaij) { 2567a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2568a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2569aa0d41d4SStefano Zampini } else { 2570a00504b5SStefano Zampini Mat work_mat; 25711cf9b237SStefano Zampini 2572a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2573aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2574a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2575aa0d41d4SStefano Zampini } 25763301b35fSStefano Zampini if (matis->A->symmetric_set) { 25773301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2578e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25793301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2580e496cd5dSStefano Zampini #endif 25813301b35fSStefano Zampini } 2582d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2583aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2584aa0d41d4SStefano Zampini } 2585aa0d41d4SStefano Zampini 2586aa0d41d4SStefano Zampini #undef __FUNCT__ 2587a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25888ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2589a64d13efSStefano Zampini { 2590a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2591a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2592d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 259353892102SStefano Zampini PetscInt *idx_R_local=NULL; 25943a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25953a50541eSStefano Zampini PetscInt vbs,bs; 25966816873aSStefano Zampini PetscBT bitmask=NULL; 2597a64d13efSStefano Zampini PetscErrorCode ierr; 2598a64d13efSStefano Zampini 2599a64d13efSStefano Zampini PetscFunctionBegin; 2600b23d619eSStefano Zampini /* 2601b23d619eSStefano Zampini No need to setup local scatters if 2602b23d619eSStefano Zampini - primal space is unchanged 2603b23d619eSStefano Zampini AND 2604b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2605b23d619eSStefano Zampini AND 2606b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2607b23d619eSStefano Zampini */ 2608b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2609f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2610f4ddd8eeSStefano Zampini } 2611f4ddd8eeSStefano Zampini /* destroy old objects */ 2612f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2613f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2614f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2615a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2616b371cd4fSStefano Zampini n_B = pcis->n_B; 2617b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2618b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 26193a50541eSStefano Zampini 2620a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 26216816873aSStefano Zampini 262253892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2623df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2624854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2625a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2626a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 26270e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2628a64d13efSStefano Zampini } 2629a64d13efSStefano Zampini 2630a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26314641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26326816873aSStefano Zampini idx_R_local[n_R++] = i; 2633a64d13efSStefano Zampini } 2634a64d13efSStefano Zampini } 2635df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2636df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26376816873aSStefano Zampini 2638df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2639df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26406816873aSStefano Zampini } 26413a50541eSStefano Zampini 26423a50541eSStefano Zampini /* Block code */ 26433a50541eSStefano Zampini vbs = 1; 26443a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26453a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26463a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26473a50541eSStefano Zampini PetscInt *vary; 2648df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2649785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26503a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2651d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2652d3df7717SStefano 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 */ 26530e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2654d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26553a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26563a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26573a50541eSStefano Zampini break; 26583a50541eSStefano Zampini } 26593a50541eSStefano Zampini } 2660d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2661d3df7717SStefano Zampini } else { 2662d3df7717SStefano Zampini /* Verify directly the R set */ 2663d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2664d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2665d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2666d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2667d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2668d3df7717SStefano Zampini break; 2669d3df7717SStefano Zampini } 2670d3df7717SStefano Zampini } 2671d3df7717SStefano Zampini } 2672d3df7717SStefano Zampini } 26733a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26743a50541eSStefano Zampini vbs = bs; 26753a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26763a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26773a50541eSStefano Zampini } 26783a50541eSStefano Zampini } 26793a50541eSStefano Zampini } 26803a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2681df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2682df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 268353892102SStefano Zampini 2684df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2685df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 268653892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2687df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 268853892102SStefano Zampini } else { 26893a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 269053892102SStefano Zampini } 2691a64d13efSStefano Zampini 2692a64d13efSStefano Zampini /* print some info if requested */ 2693a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2694a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2695a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26961575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2697a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2698a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 26994f1b2e48SStefano 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); 2700a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2701a64d13efSStefano Zampini } 2702a64d13efSStefano Zampini 2703a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2704df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 27056816873aSStefano Zampini IS is_aux1,is_aux2; 27066816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 27076816873aSStefano Zampini 27083a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2709854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2710854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2711a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 27124641a718SStefano Zampini for (i=0; i<n_D; i++) { 27134641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 27144641a718SStefano Zampini } 2715a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2716a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27174641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 27184641a718SStefano Zampini aux_array1[j++] = i; 2719a64d13efSStefano Zampini } 2720a64d13efSStefano Zampini } 2721a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2722a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2723a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 27244641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 27254641a718SStefano Zampini aux_array2[j++] = i; 2726a64d13efSStefano Zampini } 2727a64d13efSStefano Zampini } 2728a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2729a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2730a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2731a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2732a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2733a64d13efSStefano Zampini 27348eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2735785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2736a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27374641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27384641a718SStefano Zampini aux_array1[j++] = i; 2739a64d13efSStefano Zampini } 2740a64d13efSStefano Zampini } 2741a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2742a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2743a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2744a64d13efSStefano Zampini } 27454641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27463a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2747d62866d3SStefano Zampini } else { 2748df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27496816873aSStefano Zampini IS tis; 27506816873aSStefano Zampini PetscInt schur_size; 27516816873aSStefano Zampini 2752df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27536816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2754df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27556816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27566816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27576816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27586816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27596816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2760d62866d3SStefano Zampini } 2761d62866d3SStefano Zampini } 2762a64d13efSStefano Zampini PetscFunctionReturn(0); 2763a64d13efSStefano Zampini } 2764a64d13efSStefano Zampini 2765304d26faSStefano Zampini 2766304d26faSStefano Zampini #undef __FUNCT__ 2767304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2768684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2769304d26faSStefano Zampini { 2770304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2771304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2772304d26faSStefano Zampini PC pc_temp; 2773304d26faSStefano Zampini Mat A_RR; 2774f4ddd8eeSStefano Zampini MatReuse reuse; 2775304d26faSStefano Zampini PetscScalar m_one = -1.0; 2776304d26faSStefano Zampini PetscReal value; 277704708bb6SStefano Zampini PetscInt n_D,n_R; 27789577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2779304d26faSStefano Zampini PetscErrorCode ierr; 2780e604994aSStefano Zampini /* prefixes stuff */ 2781312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2782e604994aSStefano Zampini size_t len; 2783304d26faSStefano Zampini 2784304d26faSStefano Zampini PetscFunctionBegin; 2785304d26faSStefano Zampini 2786e604994aSStefano Zampini /* compute prefixes */ 2787e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2788e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2789e604994aSStefano Zampini if (!pcbddc->current_level) { 2790e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2791e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2792e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2793e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2794e604994aSStefano Zampini } else { 2795e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2796312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2797e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2798e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2799312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2800312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 280134d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 280234d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2803e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2804e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2805e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2806e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2807e604994aSStefano Zampini } 2808e604994aSStefano Zampini 2809304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2810684f6988SStefano Zampini if (dirichlet) { 2811d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2812450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2813450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2814450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2815450f8f5eSStefano Zampini } 2816450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2817a3df083aSStefano Zampini Mat A_IIn; 2818a3df083aSStefano Zampini 2819a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2820a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2821a3df083aSStefano Zampini pcis->A_II = A_IIn; 2822a3df083aSStefano Zampini } 2823450f8f5eSStefano Zampini } 28243301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28253301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2826964fefecSStefano Zampini } 2827ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2828964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2829304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2830304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2831304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2832304d26faSStefano Zampini /* default */ 2833304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2834e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28359577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2836304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28379577ea80SStefano Zampini if (issbaij) { 28389577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28399577ea80SStefano Zampini } else { 2840304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28419577ea80SStefano Zampini } 2842304d26faSStefano Zampini /* Allow user's customization */ 2843304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2844304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2845304d26faSStefano Zampini } 2846d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2847df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2848df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2849d62866d3SStefano Zampini 2850df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2851d5574798SStefano Zampini } 2852304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2853304d26faSStefano Zampini if (!n_D) { 2854304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2855304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2856304d26faSStefano Zampini } 2857304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2858304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2859304d26faSStefano Zampini /* set ksp_D into pcis data */ 2860304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2861304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2862304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2863684f6988SStefano Zampini } 2864304d26faSStefano Zampini 2865304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2866684f6988SStefano Zampini A_RR = 0; 2867684f6988SStefano Zampini if (neumann) { 2868d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 286904708bb6SStefano Zampini PetscInt ibs,mbs; 287004708bb6SStefano Zampini PetscBool issbaij; 287104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2872f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28738ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2874f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2875f4ddd8eeSStefano Zampini PetscInt nn_R; 287681d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2877f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2878f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2879f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2880f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2881f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2882f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2883f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2884727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2885f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2886f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2887f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2888f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2889f4ddd8eeSStefano Zampini } 2890f4ddd8eeSStefano Zampini } 2891f4ddd8eeSStefano Zampini /* last check */ 2892d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2893f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2894f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2895f4ddd8eeSStefano Zampini } 2896f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2897f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2898f4ddd8eeSStefano Zampini } 2899a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 2900af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2901af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 290204708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 290304708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 290404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 290504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 290604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2907af732b37SStefano Zampini } else { 2908511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 29096816873aSStefano Zampini } 291004708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 291104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 291404708bb6SStefano Zampini } else { 2915511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 291604708bb6SStefano Zampini } 291704708bb6SStefano Zampini } 2918a00504b5SStefano Zampini /* extract A_RR */ 2919a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 2920a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2921a00504b5SStefano Zampini 2922a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 292316e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2924a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 292516e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 292616e386b8SStefano Zampini } else { 2927a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 2928a00504b5SStefano Zampini } 2929a00504b5SStefano Zampini } else { 2930a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2931a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 2932a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2933a00504b5SStefano Zampini } 2934a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 2935f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 293616e386b8SStefano Zampini } 29373301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29383301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29396816873aSStefano Zampini } 2940f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2941304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2942304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2943304d26faSStefano Zampini /* default */ 2944304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2945e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2946304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29479577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29489577ea80SStefano Zampini if (issbaij) { 29499577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29509577ea80SStefano Zampini } else { 2951304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29529577ea80SStefano Zampini } 2953304d26faSStefano Zampini /* Allow user's customization */ 2954304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2955304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2956304d26faSStefano Zampini } 2957304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2958304d26faSStefano Zampini if (!n_R) { 2959304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2960304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2961304d26faSStefano Zampini } 29625cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2963df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2964df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2965df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2966d62866d3SStefano Zampini 2967df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2968d62866d3SStefano Zampini } 2969304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2970304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2971684f6988SStefano Zampini } 2972304d26faSStefano Zampini 2973304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29740fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2975684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2976684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29771575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2978684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2979684f6988SStefano Zampini } 2980684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29810fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29820fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29830fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29840fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29850fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2986304d26faSStefano Zampini /* need to be adapted? */ 2987b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2988b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2989b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2990304d26faSStefano Zampini /* print info */ 2991304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2992e604994aSStefano 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); 2993304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2994304d26faSStefano Zampini } 2995b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2996298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2997304d26faSStefano Zampini } 2998684f6988SStefano Zampini } 2999684f6988SStefano Zampini if (neumann) { /* Neumann */ 30000fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 30010fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 30020fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 30030fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 30040fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3005304d26faSStefano Zampini /* need to be adapted? */ 3006b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 3007b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3008304d26faSStefano Zampini /* print info */ 3009304d26faSStefano Zampini if (pcbddc->dbg_flag) { 3010e604994aSStefano 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); 3011304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3012304d26faSStefano Zampini } 3013b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 3014298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 3015304d26faSStefano Zampini } 30160fccc4e9SStefano Zampini } 3017684f6988SStefano Zampini } 30185cbda25cSStefano Zampini /* free Neumann problem's matrix */ 30195cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3020304d26faSStefano Zampini PetscFunctionReturn(0); 3021304d26faSStefano Zampini } 3022304d26faSStefano Zampini 3023304d26faSStefano Zampini #undef __FUNCT__ 3024ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 302580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3026674ae819SStefano Zampini { 3027674ae819SStefano Zampini PetscErrorCode ierr; 3028674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3029be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3030674ae819SStefano Zampini 3031674ae819SStefano Zampini PetscFunctionBegin; 3032df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 303380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 303420c7b377SStefano Zampini } 303580677318SStefano Zampini if (!pcbddc->switch_static) { 303680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 303780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 303880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 303920c7b377SStefano Zampini } 3040df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 304180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304320c7b377SStefano Zampini } else { 3044df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3045be83ff47SStefano Zampini 3046df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3047df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 304820c7b377SStefano Zampini } 3049be83ff47SStefano Zampini } else { 305080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 305580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 305680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 305780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3059674ae819SStefano Zampini } 3060674ae819SStefano Zampini } 306172b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 306280677318SStefano Zampini if (applytranspose) { 306380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 306480677318SStefano Zampini } else { 306580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 306680677318SStefano Zampini } 3067be83ff47SStefano Zampini } else { 3068df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3069be83ff47SStefano Zampini 3070be83ff47SStefano Zampini if (applytranspose) { 3071df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3072be83ff47SStefano Zampini } else { 3073df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3074be83ff47SStefano Zampini } 3075be83ff47SStefano Zampini } 307680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 307780677318SStefano Zampini if (!pcbddc->switch_static) { 3078df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 307980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3081be83ff47SStefano Zampini } else { 3082df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3083be83ff47SStefano Zampini 3084df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3085df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3086be83ff47SStefano Zampini } 308780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 308880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 308980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 309080677318SStefano Zampini } 309180677318SStefano Zampini } else { 309280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 309780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 309880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 309980677318SStefano Zampini } 310080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3104674ae819SStefano Zampini } 3105674ae819SStefano Zampini PetscFunctionReturn(0); 3106674ae819SStefano Zampini } 3107674ae819SStefano Zampini 3108dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3109674ae819SStefano Zampini #undef __FUNCT__ 3110674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3111dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3112674ae819SStefano Zampini { 3113674ae819SStefano Zampini PetscErrorCode ierr; 3114674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3115674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3116674ae819SStefano Zampini const PetscScalar zero = 0.0; 3117674ae819SStefano Zampini 3118674ae819SStefano Zampini PetscFunctionBegin; 3119dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 31204fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3121dc359a40SStefano Zampini if (applytranspose) { 3122674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 31238eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3124dc359a40SStefano Zampini } else { 3125674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3126674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 312715aaf578SStefano Zampini } 31284fee134fSStefano Zampini } else { 31294fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 31304fee134fSStefano Zampini } 3131efc2fbd9SStefano Zampini 3132efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31334f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3134efc2fbd9SStefano Zampini PetscScalar *array; 31354f1b2e48SStefano Zampini PetscInt j; 3136efc2fbd9SStefano Zampini 3137efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31384f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3139efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3140efc2fbd9SStefano Zampini } 3141efc2fbd9SStefano Zampini 314212edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 314312edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 314412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 314512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 314612edc857SStefano Zampini 31479f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 314812edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 314912edc857SStefano Zampini if (pcbddc->coarse_ksp) { 315051694757SStefano Zampini Mat coarse_mat; 3151964fefecSStefano Zampini Vec rhs,sol; 315251694757SStefano Zampini MatNullSpace nullsp; 3153964fefecSStefano Zampini 3154964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3155964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 315651694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 315751694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 315851694757SStefano Zampini if (nullsp) { 315951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 316051694757SStefano Zampini } 316112edc857SStefano Zampini if (applytranspose) { 3162964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 316312edc857SStefano Zampini } else { 3164964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 316512edc857SStefano Zampini } 316651694757SStefano Zampini if (nullsp) { 316751694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 316851694757SStefano Zampini } 316912edc857SStefano Zampini } 3170674ae819SStefano Zampini 3171674ae819SStefano Zampini /* Local solution on R nodes */ 31724fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 317380677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31749f00e9b4SStefano Zampini } 31759f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31769f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 317712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3178674ae819SStefano Zampini 31794fee134fSStefano Zampini /* Sum contributions from the two levels */ 31804fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3181dc359a40SStefano Zampini if (applytranspose) { 3182dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3183dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3184dc359a40SStefano Zampini } else { 3185674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31868eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3187dc359a40SStefano Zampini } 3188efc2fbd9SStefano Zampini /* store p0 */ 31894f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3190efc2fbd9SStefano Zampini PetscScalar *array; 31914f1b2e48SStefano Zampini PetscInt j; 3192efc2fbd9SStefano Zampini 3193efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31944f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3195efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3196efc2fbd9SStefano Zampini } 31974fee134fSStefano Zampini } else { /* expand the coarse solution */ 31984fee134fSStefano Zampini if (applytranspose) { 31994fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32004fee134fSStefano Zampini } else { 32014fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32024fee134fSStefano Zampini } 32034fee134fSStefano Zampini } 3204674ae819SStefano Zampini PetscFunctionReturn(0); 3205674ae819SStefano Zampini } 3206674ae819SStefano Zampini 3207674ae819SStefano Zampini #undef __FUNCT__ 3208674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 320912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3210674ae819SStefano Zampini { 3211674ae819SStefano Zampini PetscErrorCode ierr; 3212674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 321358da7f69SStefano Zampini PetscScalar *array; 321412edc857SStefano Zampini Vec from,to; 3215674ae819SStefano Zampini 3216674ae819SStefano Zampini PetscFunctionBegin; 321712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 321812edc857SStefano Zampini from = pcbddc->coarse_vec; 321912edc857SStefano Zampini to = pcbddc->vec1_P; 322012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 322112edc857SStefano Zampini Vec tvec; 322258da7f69SStefano Zampini 322358da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 322458da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 322512edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 322658da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 322758da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 322858da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 322912edc857SStefano Zampini } 323012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 323112edc857SStefano Zampini from = pcbddc->vec1_P; 323212edc857SStefano Zampini to = pcbddc->coarse_vec; 323312edc857SStefano Zampini } 323412edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3235674ae819SStefano Zampini PetscFunctionReturn(0); 3236674ae819SStefano Zampini } 3237674ae819SStefano Zampini 3238674ae819SStefano Zampini #undef __FUNCT__ 3239674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 324012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3241674ae819SStefano Zampini { 3242674ae819SStefano Zampini PetscErrorCode ierr; 3243674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 324458da7f69SStefano Zampini PetscScalar *array; 324512edc857SStefano Zampini Vec from,to; 3246674ae819SStefano Zampini 3247674ae819SStefano Zampini PetscFunctionBegin; 324812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 324912edc857SStefano Zampini from = pcbddc->coarse_vec; 325012edc857SStefano Zampini to = pcbddc->vec1_P; 325112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 325212edc857SStefano Zampini from = pcbddc->vec1_P; 325312edc857SStefano Zampini to = pcbddc->coarse_vec; 325412edc857SStefano Zampini } 325512edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 325612edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 325712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 325812edc857SStefano Zampini Vec tvec; 325958da7f69SStefano Zampini 326012edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 326158da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 326258da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 326358da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 326458da7f69SStefano Zampini } 326558da7f69SStefano Zampini } else { 326658da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 326758da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 326812edc857SStefano Zampini } 326912edc857SStefano Zampini } 3270674ae819SStefano Zampini PetscFunctionReturn(0); 3271674ae819SStefano Zampini } 3272674ae819SStefano Zampini 3273984c4197SStefano Zampini /* uncomment for testing purposes */ 3274984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3275674ae819SStefano Zampini #undef __FUNCT__ 3276674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3277674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3278674ae819SStefano Zampini { 3279674ae819SStefano Zampini PetscErrorCode ierr; 3280674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3281674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3282674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3283984c4197SStefano Zampini /* one and zero */ 3284984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3285984c4197SStefano Zampini /* space to store constraints and their local indices */ 32869162d606SStefano Zampini PetscScalar *constraints_data; 32879162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32889162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32899162d606SStefano Zampini PetscInt *constraints_n; 3290984c4197SStefano Zampini /* iterators */ 3291b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3292984c4197SStefano Zampini /* BLAS integers */ 3293e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3294e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3295c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3296727cdba6SStefano Zampini /* reuse */ 32970e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 32980e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3299984c4197SStefano Zampini /* change of basis */ 3300b3d85658SStefano Zampini PetscBool qr_needed; 33019162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3302984c4197SStefano Zampini /* auxiliary stuff */ 330364efe560SStefano Zampini PetscInt *nnz,*is_indices; 33048a0068c3SStefano Zampini PetscInt ncc; 3305984c4197SStefano Zampini /* some quantities */ 330645a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3307a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3308984c4197SStefano Zampini 3309674ae819SStefano Zampini PetscFunctionBegin; 33108e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 33118e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 33128e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 331316909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3314088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3315088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 33160e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 33170e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 33180e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33190e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33200e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3321088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3322cf5a6209SStefano Zampini 3323cf5a6209SStefano Zampini /* print some info */ 3324cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3325cf5a6209SStefano Zampini IS vertices; 3326cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3327cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3328cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3329cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 33301575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3331cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3332cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3333fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3334fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3335cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33361575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3337cf5a6209SStefano Zampini } 3338cf5a6209SStefano Zampini 3339cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33409162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3341cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3342cf5a6209SStefano Zampini const Vec *nearnullvecs; 3343cf5a6209SStefano Zampini Vec *localnearnullsp; 3344cf5a6209SStefano Zampini PetscScalar *array; 3345cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3346cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3347674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3348b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3349674ae819SStefano Zampini PetscScalar *work; 3350674ae819SStefano Zampini PetscReal *singular_vals; 3351674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3352674ae819SStefano Zampini PetscReal *rwork; 3353674ae819SStefano Zampini #endif 3354674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3355674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3356674ae819SStefano Zampini #else 3357964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3358964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3359674ae819SStefano Zampini #endif 3360674ae819SStefano Zampini 3361674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3362d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3363d06fc5fdSStefano Zampini /* free unneeded index sets */ 3364d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3365d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3366674ae819SStefano Zampini } 3367d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3368d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3369d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3370d06fc5fdSStefano Zampini } 3371d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3372d06fc5fdSStefano Zampini n_ISForEdges = 0; 3373d06fc5fdSStefano Zampini } 3374d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3375d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3376d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3377d06fc5fdSStefano Zampini } 3378d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3379d06fc5fdSStefano Zampini n_ISForFaces = 0; 3380d06fc5fdSStefano Zampini } 338170022509SStefano Zampini 338270022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 338370022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 338470022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 338570022509SStefano Zampini if (pcbddc->NullSpace) { 338670022509SStefano Zampini PetscBool tbool[2],gbool[2]; 338770022509SStefano Zampini 338870022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3389b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3390d06fc5fdSStefano Zampini if (!ISForEdges) { 3391d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3392d06fc5fdSStefano Zampini } 3393b8ffe317SStefano Zampini } 3394d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3395d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3396b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3397d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3398d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 339998a51de6SStefano Zampini } 340070022509SStefano Zampini #endif 340108122e43SStefano Zampini 3402674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3403674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3404674ae819SStefano Zampini if (nearnullsp) { 3405674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3406f4ddd8eeSStefano Zampini /* remove any stored info */ 3407f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3408f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3409f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3410f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3411f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3412473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3413f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3414f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3415f4ddd8eeSStefano Zampini } 3416984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3417984c4197SStefano Zampini nnsp_size = 0; 3418674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3419674ae819SStefano Zampini } 3420984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3421984c4197SStefano Zampini max_constraints = nnsp_size; 3422984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3423984c4197SStefano Zampini 3424674ae819SStefano Zampini /* 3425674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 34269162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 34279162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 34289162d606SStefano Zampini There can be multiple constraints per connected component 3429674ae819SStefano Zampini */ 3430674ae819SStefano Zampini n_vertices = 0; 3431674ae819SStefano Zampini if (ISForVertices) { 3432674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3433674ae819SStefano Zampini } 34349162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34359162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34369162d606SStefano Zampini 34379162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34389162d606SStefano Zampini total_counts *= max_constraints; 3439674ae819SStefano Zampini total_counts += n_vertices; 34404641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34419162d606SStefano Zampini 3442674ae819SStefano Zampini total_counts = 0; 3443674ae819SStefano Zampini max_size_of_constraint = 0; 3444674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34459162d606SStefano Zampini IS used_is; 3446674ae819SStefano Zampini if (i<n_ISForEdges) { 34479162d606SStefano Zampini used_is = ISForEdges[i]; 3448674ae819SStefano Zampini } else { 34499162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3450674ae819SStefano Zampini } 34519162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3452674ae819SStefano Zampini total_counts += j; 3453674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3454674ae819SStefano Zampini } 34559162d606SStefano 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); 34569162d606SStefano Zampini 3457984c4197SStefano Zampini /* get local part of global near null space vectors */ 3458785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3459984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3460984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3461e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3462e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3463984c4197SStefano Zampini } 3464674ae819SStefano Zampini 3465242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3466242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3467a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3468242a89d7SStefano Zampini 3469984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3470a773dcb8SStefano Zampini if (!skip_lapack) { 3471674ae819SStefano Zampini PetscScalar temp_work; 3472911cabfeSStefano Zampini 3473674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3474984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3475785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3476785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3477785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3478674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3479785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3480674ae819SStefano Zampini #endif 3481674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3482c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3483c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3484674ae819SStefano Zampini lwork = -1; 3485674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3486674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3487c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3488674ae819SStefano Zampini #else 3489c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3490674ae819SStefano Zampini #endif 3491674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3492984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3493674ae819SStefano Zampini #else /* on missing GESVD */ 3494674ae819SStefano Zampini /* SVD */ 3495674ae819SStefano Zampini PetscInt max_n,min_n; 3496674ae819SStefano Zampini max_n = max_size_of_constraint; 3497984c4197SStefano Zampini min_n = max_constraints; 3498984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3499674ae819SStefano Zampini min_n = max_size_of_constraint; 3500984c4197SStefano Zampini max_n = max_constraints; 3501674ae819SStefano Zampini } 3502785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3503674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3504785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3505674ae819SStefano Zampini #endif 3506674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3507674ae819SStefano Zampini lwork = -1; 3508e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3509e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3510b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3511674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3512674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35139162d606SStefano 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)); 3514674ae819SStefano Zampini #else 35159162d606SStefano 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)); 3516674ae819SStefano Zampini #endif 3517674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3518984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3519984c4197SStefano Zampini #endif /* on missing GESVD */ 3520674ae819SStefano Zampini /* Allocate optimal workspace */ 3521674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3522854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3523674ae819SStefano Zampini } 3524674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3525674ae819SStefano Zampini total_counts = 0; 35269162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 35279162d606SStefano Zampini constraints_data_ptr[0] = 0; 3528674ae819SStefano Zampini /* vertices */ 35299162d606SStefano Zampini if (n_vertices) { 3530674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 35319162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3532674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35339162d606SStefano Zampini constraints_n[total_counts] = 1; 35349162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35359162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35369162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3537674ae819SStefano Zampini total_counts++; 3538674ae819SStefano Zampini } 3539674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3540674ae819SStefano Zampini n_vertices = total_counts; 3541674ae819SStefano Zampini } 3542984c4197SStefano Zampini 3543674ae819SStefano Zampini /* edges and faces */ 35449162d606SStefano Zampini total_counts_cc = total_counts; 3545911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35469162d606SStefano Zampini IS used_is; 35479162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35489162d606SStefano Zampini 3549911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35509162d606SStefano Zampini used_is = ISForEdges[ncc]; 3551984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3552674ae819SStefano Zampini } else { 35539162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3554984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3555674ae819SStefano Zampini } 3556674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35579162d606SStefano Zampini 35589162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35599162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3560984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3561984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3562674ae819SStefano Zampini if (nnsp_has_cnst) { 35635b08dc53SStefano Zampini PetscScalar quad_value; 35649162d606SStefano Zampini 35659162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35669162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35679162d606SStefano Zampini 3568a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3569674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3570a773dcb8SStefano Zampini } else { 3571a773dcb8SStefano Zampini quad_value = 1.0; 3572a773dcb8SStefano Zampini } 3573674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35749162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3575674ae819SStefano Zampini } 35769162d606SStefano Zampini temp_constraints++; 3577674ae819SStefano Zampini total_counts++; 3578674ae819SStefano Zampini } 3579674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3580984c4197SStefano Zampini PetscReal real_value; 35819162d606SStefano Zampini PetscScalar *ptr_to_data; 35829162d606SStefano Zampini 3583984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35849162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3585674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35869162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3587674ae819SStefano Zampini } 3588984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3589984c4197SStefano Zampini /* check if array is null on the connected component */ 3590e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35919162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35925b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3593674ae819SStefano Zampini temp_constraints++; 3594674ae819SStefano Zampini total_counts++; 35959162d606SStefano Zampini if (!idxs_copied) { 35969162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35979162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3598674ae819SStefano Zampini } 3599674ae819SStefano Zampini } 36009162d606SStefano Zampini } 36019162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 360245a1bb75SStefano Zampini valid_constraints = temp_constraints; 3603eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3604a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 36059162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 36069162d606SStefano Zampini 36079162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3608a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 36099162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3610a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 36119162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3612a773dcb8SStefano Zampini } else { /* perform SVD */ 3613984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 36149162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3615674ae819SStefano Zampini 3616674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3617984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3618984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3619984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3620984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3621984c4197SStefano Zampini from that computed using LAPACKgesvd 3622984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3623984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3624984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3625674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3626e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3627984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3628674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3629674ae819SStefano Zampini for (k=0;k<j+1;k++) { 36309162d606SStefano 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)); 3631674ae819SStefano Zampini } 3632674ae819SStefano Zampini } 3633e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3634e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3635e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3636674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3637c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3638674ae819SStefano Zampini #else 3639c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3640674ae819SStefano Zampini #endif 3641674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3642984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3643984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3644674ae819SStefano Zampini j = 0; 3645984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3646674ae819SStefano Zampini total_counts = total_counts-j; 364745a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3648e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3649c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3650c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3651c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3652c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3653c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3654c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3655674ae819SStefano Zampini if (j<temp_constraints) { 3656984c4197SStefano Zampini PetscInt ii; 3657984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3658674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36599162d606SStefano 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)); 3660674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3661984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3662674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36639162d606SStefano 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]; 3664674ae819SStefano Zampini } 3665674ae819SStefano Zampini } 3666674ae819SStefano Zampini } 3667674ae819SStefano Zampini #else /* on missing GESVD */ 3668e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3669e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3670b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3671674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3672674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36739162d606SStefano 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)); 3674674ae819SStefano Zampini #else 36759162d606SStefano 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)); 3676674ae819SStefano Zampini #endif 3677984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3678674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3679984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3680e310c8b4SStefano Zampini k = temp_constraints; 3681e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3682674ae819SStefano Zampini j = 0; 3683e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 368445a1bb75SStefano Zampini valid_constraints = k-j; 3685911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3686984c4197SStefano Zampini #endif /* on missing GESVD */ 3687674ae819SStefano Zampini } 3688a773dcb8SStefano Zampini } 36899162d606SStefano Zampini /* update pointers information */ 36909162d606SStefano Zampini if (valid_constraints) { 36919162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36929162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36939162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36949162d606SStefano Zampini /* set change_of_basis flag */ 369545a1bb75SStefano Zampini if (boolforchange) { 3696b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 36979162d606SStefano Zampini } 3698b3d85658SStefano Zampini total_counts_cc++; 369945a1bb75SStefano Zampini } 370045a1bb75SStefano Zampini } 3701984c4197SStefano Zampini /* free workspace */ 37028f1c130eSStefano Zampini if (!skip_lapack) { 3703984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3704984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3705984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3706984c4197SStefano Zampini #endif 3707984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3708984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3709984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3710984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3711984c4197SStefano Zampini #endif 3712984c4197SStefano Zampini } 3713984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3714984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3715984c4197SStefano Zampini } 3716984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3717cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3718cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3719cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3720cf5a6209SStefano Zampini } 3721cf5a6209SStefano Zampini if (n_ISForFaces) { 3722cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3723cf5a6209SStefano Zampini } 3724cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3725cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3726cf5a6209SStefano Zampini } 3727cf5a6209SStefano Zampini if (n_ISForEdges) { 3728cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3729cf5a6209SStefano Zampini } 3730cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 373108122e43SStefano Zampini } else { 373208122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3733984c4197SStefano Zampini 373408122e43SStefano Zampini total_counts = 0; 373508122e43SStefano Zampini n_vertices = 0; 3736d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3737d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 373808122e43SStefano Zampini } 373908122e43SStefano Zampini max_constraints = 0; 37409162d606SStefano Zampini total_counts_cc = 0; 374108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 374208122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37439162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 374408122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 374508122e43SStefano Zampini } 37469162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37479162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37489162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37499162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 375074d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37519162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37529162d606SStefano Zampini total_counts_cc = 0; 37539162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37549162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37559162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 375608122e43SStefano Zampini } 375708122e43SStefano Zampini } 37589162d606SStefano Zampini #if 0 37599162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37609162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37619162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37629162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37639162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37649162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37659162d606SStefano Zampini } 37669162d606SStefano Zampini printf("\n"); 37679162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37689162d606SStefano Zampini } 37691b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37708bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37711b968477SStefano Zampini } 37721b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37738bec7fa6SStefano 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]); 37741b968477SStefano Zampini } 377508122e43SStefano Zampini #endif 377608122e43SStefano Zampini 37778bec7fa6SStefano Zampini max_size_of_constraint = 0; 37789162d606SStefano 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]); 37799162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 378008122e43SStefano Zampini /* Change of basis */ 3781b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 378208122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 378308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 378408122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3785b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 378608122e43SStefano Zampini } 378708122e43SStefano Zampini } 378808122e43SStefano Zampini } 378908122e43SStefano Zampini } 3790984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37914f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 379208122e43SStefano Zampini 37939162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37949162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 37956c4ed002SBarry 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); 3796674ae819SStefano Zampini 3797674ae819SStefano Zampini /* Create constraint matrix */ 3798674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 379916f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3800984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3801984c4197SStefano Zampini 3802984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3803a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3804a717540cSStefano Zampini qr_needed = PETSC_FALSE; 380574d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3806984c4197SStefano Zampini total_primal_vertices=0; 3807b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 38089162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 38099162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 381072b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 38119162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3812b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 381364efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 38149162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38159162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3816a717540cSStefano Zampini } 3817b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 381874d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3819a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3820a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3821a717540cSStefano Zampini } 3822fa434743SStefano Zampini } else { 3823b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3824fa434743SStefano Zampini } 3825a717540cSStefano Zampini } 3826b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3827b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3828674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 382970022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3830b3d85658SStefano Zampini 38314f1b2e48SStefano 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); 38320e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38330e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3834984c4197SStefano Zampini 3835984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 383674d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3837785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3838984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 383974d5cdf7SStefano Zampini 3840984c4197SStefano Zampini j = total_primal_vertices; 384174d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3842b3d85658SStefano Zampini cum = total_primal_vertices; 38439162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38444641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3845b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3846b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3847b3d85658SStefano Zampini cum++; 38489162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 384974d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 385074d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 385174d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 385274d5cdf7SStefano Zampini } 38539162d606SStefano Zampini j += constraints_n[i]; 3854674ae819SStefano Zampini } 3855674ae819SStefano Zampini } 3856674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3857674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3858088faed8SStefano Zampini 3859674ae819SStefano Zampini /* set values in constraint matrix */ 3860984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38610e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3862674ae819SStefano Zampini } 3863984c4197SStefano Zampini total_counts = total_primal_vertices; 38649162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38654641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38669162d606SStefano Zampini PetscInt *cols; 38679162d606SStefano Zampini 38689162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38699162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38709162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38719162d606SStefano Zampini PetscInt row = total_counts+k; 38729162d606SStefano Zampini PetscScalar *vals; 38739162d606SStefano Zampini 38749162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38759162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38769162d606SStefano Zampini } 38779162d606SStefano Zampini total_counts += constraints_n[i]; 3878674ae819SStefano Zampini } 3879674ae819SStefano Zampini } 3880674ae819SStefano Zampini /* assembling */ 3881674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3882674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3883088faed8SStefano Zampini 3884984c4197SStefano Zampini /* 38856a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3886984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3887f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 3888984c4197SStefano Zampini */ 3889674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3890674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3891026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3892984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3893984c4197SStefano Zampini /* working stuff for GEQRF */ 389481d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3895984c4197SStefano Zampini PetscBLASInt lqr_work; 3896984c4197SStefano Zampini /* working stuff for UNGQR */ 3897984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3898984c4197SStefano Zampini PetscBLASInt lgqr_work; 3899984c4197SStefano Zampini /* working stuff for TRTRS */ 3900984c4197SStefano Zampini PetscScalar *trs_rhs; 39013f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3902984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3903984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3904984c4197SStefano Zampini PetscScalar *start_vals; 3905984c4197SStefano Zampini /* working stuff for values insertion */ 39064641a718SStefano Zampini PetscBT is_primal; 390764efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3908906d46d4SStefano Zampini /* matrix sizes */ 3909906d46d4SStefano Zampini PetscInt global_size,local_size; 3910906d46d4SStefano Zampini /* temporary change of basis */ 3911906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3912cf5a6209SStefano Zampini /* extra space for debugging */ 3913cf5a6209SStefano Zampini PetscScalar *dbg_work; 3914984c4197SStefano Zampini 3915906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3916906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 391716f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3918bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3919906d46d4SStefano Zampini /* nonzeros for local mat */ 3920bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 39211dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3922bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 39231dd7afcfSStefano Zampini } else { 39241dd7afcfSStefano Zampini const PetscInt *ii; 39251dd7afcfSStefano Zampini PetscInt n; 39261dd7afcfSStefano Zampini PetscBool flg_row; 39271dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39281dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 39291dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39301dd7afcfSStefano Zampini } 39319162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3932a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39339162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3934a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39359162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3936a717540cSStefano Zampini } else { 39379162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39389162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3939a717540cSStefano Zampini } 3940a717540cSStefano Zampini } 3941a717540cSStefano Zampini } 3942906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3943bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 39441dd7afcfSStefano Zampini /* Set interior change in the matrix */ 39451dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3946bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3947906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3948a717540cSStefano Zampini } 39491dd7afcfSStefano Zampini } else { 39501dd7afcfSStefano Zampini const PetscInt *ii,*jj; 39511dd7afcfSStefano Zampini PetscScalar *aa; 39521dd7afcfSStefano Zampini PetscInt n; 39531dd7afcfSStefano Zampini PetscBool flg_row; 39541dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39551dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39561dd7afcfSStefano Zampini for (i=0;i<n;i++) { 39571dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 39581dd7afcfSStefano Zampini } 39591dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39601dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39611dd7afcfSStefano Zampini } 3962a717540cSStefano Zampini 3963a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3964a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3965a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3966a717540cSStefano Zampini } 3967a717540cSStefano Zampini 3968a717540cSStefano Zampini 3969a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3970a717540cSStefano Zampini /* 3971a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3972a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3973a717540cSStefano Zampini 3974a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3975a717540cSStefano Zampini 3976a6b551f4SStefano 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) 3977a6b551f4SStefano Zampini 3978a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3979a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3980a717540cSStefano Zampini | ... | 3981a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3982a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3983a717540cSStefano Zampini 3984a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3985a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3986a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3987a6b551f4SStefano Zampini 3988a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3989a717540cSStefano Zampini */ 3990a717540cSStefano Zampini if (qr_needed) { 3991984c4197SStefano Zampini /* space to store Q */ 3992854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3993984c4197SStefano Zampini /* first we issue queries for optimal work */ 39943f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39953f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 39963f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3997984c4197SStefano Zampini lqr_work = -1; 39983f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3999984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4000984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4001785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4002984c4197SStefano Zampini lgqr_work = -1; 40033f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 40043f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 40053f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 40063f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40073f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 40083f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4009984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4010984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4011785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4012984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4013785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4014984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4015785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4016a717540cSStefano Zampini /* allocating workspace for check */ 4017a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4018cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4019a717540cSStefano Zampini } 4020a717540cSStefano Zampini } 4021984c4197SStefano Zampini /* array to store whether a node is primal or not */ 40224641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4023473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 40240e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 40256c4ed002SBarry 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); 402639e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 402739e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 402839e2fb2aSStefano Zampini } 402939e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4030984c4197SStefano Zampini 4031a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 40329162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40339162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40344641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4035984c4197SStefano Zampini /* get constraint info */ 40369162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4037984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4038984c4197SStefano Zampini 4039984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40409162d606SStefano 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); 4041674ae819SStefano Zampini } 4042984c4197SStefano Zampini 4043fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4044a717540cSStefano Zampini 4045a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4046a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40479162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4048a717540cSStefano Zampini } 4049984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40509162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4051984c4197SStefano Zampini 4052984c4197SStefano Zampini /* compute QR decomposition of constraints */ 40533f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40543f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40553f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4056674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40573f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4058984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4059674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4060984c4197SStefano Zampini 4061984c4197SStefano Zampini /* explictly compute R^-T */ 4062984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4063984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40643f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40653f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40663f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40673f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4068984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40693f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4070984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4071984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4072984c4197SStefano Zampini 4073a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40743f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40753f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40763f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40773f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4078984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40793f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4080984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4081984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4082984c4197SStefano Zampini 4083984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4084984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4085984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40863f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40873f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40883f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40893f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4092984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40939162d606SStefano 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)); 4094984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40959162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4096984c4197SStefano Zampini 4097984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 40989162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4099984c4197SStefano Zampini /* insert cols for primal dofs */ 4100984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4101984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 41029162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4103906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4104984c4197SStefano Zampini } 4105984c4197SStefano Zampini /* insert cols for dual dofs */ 4106984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 41079162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4108984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 41099162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4110906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4111984c4197SStefano Zampini j++; 4112674ae819SStefano Zampini } 4113674ae819SStefano Zampini } 4114984c4197SStefano Zampini 4115984c4197SStefano Zampini /* check change of basis */ 4116984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4117984c4197SStefano Zampini PetscInt ii,jj; 4118984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4119c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4120c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4121c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4122c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4123c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4124c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4125984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4126cf5a6209SStefano 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)); 4127984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4128984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4129984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4130cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4131cf5a6209SStefano 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; 4132674ae819SStefano Zampini } 4133674ae819SStefano Zampini } 4134984c4197SStefano Zampini if (!valid_qr) { 413522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4136984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4137984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4138cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4139cf5a6209SStefano 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])); 4140674ae819SStefano Zampini } 4141cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4142cf5a6209SStefano 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])); 4143984c4197SStefano Zampini } 4144984c4197SStefano Zampini } 4145984c4197SStefano Zampini } 4146674ae819SStefano Zampini } else { 414722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4148674ae819SStefano Zampini } 4149674ae819SStefano Zampini } 4150a717540cSStefano Zampini } else { /* simple transformation block */ 4151a717540cSStefano Zampini PetscInt row,col; 4152a6b551f4SStefano Zampini PetscScalar val,norm; 4153a6b551f4SStefano Zampini 4154a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41559162d606SStefano 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)); 4156a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 41579162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 41589162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4159bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 41609162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4161906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41629162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4163a717540cSStefano Zampini } else { 4164a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41659162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4166a717540cSStefano Zampini if (row != col) { 41679162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4168a717540cSStefano Zampini } else { 41699162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4170a717540cSStefano Zampini } 4171906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4172a717540cSStefano Zampini } 4173a717540cSStefano Zampini } 4174a717540cSStefano Zampini } 417598a51de6SStefano Zampini if (pcbddc->dbg_flag) { 417622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4177a717540cSStefano Zampini } 4178674ae819SStefano Zampini } 4179984c4197SStefano Zampini } else { 4180984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41819162d606SStefano 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); 4182674ae819SStefano Zampini } 4183674ae819SStefano Zampini } 4184674ae819SStefano Zampini } 4185a717540cSStefano Zampini 4186a717540cSStefano Zampini /* free workspace */ 4187a717540cSStefano Zampini if (qr_needed) { 4188984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4189cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4190984c4197SStefano Zampini } 4191984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4192984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4193984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4194984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4195984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4196674ae819SStefano Zampini } 4197a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4198906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4199906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4200906d46d4SStefano Zampini 4201906d46d4SStefano Zampini /* assembling of global change of variable */ 420288c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4203bbb9e6c6SStefano Zampini Mat tmat; 420416f15bc4SStefano Zampini PetscInt bs; 420516f15bc4SStefano Zampini 4206906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4207906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4208bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4209bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4210bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4211bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 421216f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 421316f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4214906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4215bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4216bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4217bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4218bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4219bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4220e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4221e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4222bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4223bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 422488c03ad3SStefano Zampini 4225906d46d4SStefano Zampini /* check */ 4226906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4227906d46d4SStefano Zampini PetscReal error; 4228906d46d4SStefano Zampini Vec x,x_change; 4229906d46d4SStefano Zampini 4230906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4231906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4232906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4233906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4234e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4235e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4236bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4237e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4238e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4239906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4240906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4241906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4242906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4243bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4244906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4245906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4246906d46d4SStefano Zampini } 4247b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4248b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4249b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4250bf3a8328SStefano Zampini 4251bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4252bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4253bf3a8328SStefano Zampini } 4254b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4255ac632422SStefano Zampini Mat S_new,tmat; 4256bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4257bbb9e6c6SStefano Zampini 4258bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 42596816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4260bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4261bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4262bf3a8328SStefano Zampini IS is_V; 4263b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4264b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4265b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4266b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4267b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4268bf3a8328SStefano Zampini } 4269bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4270ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4271b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4272ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4273bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4274bf3a8328SStefano Zampini const PetscScalar *array; 4275bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4276bf3a8328SStefano Zampini PetscInt i,n_V; 4277bf3a8328SStefano Zampini 4278b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4279b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4280b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4281b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4282b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4283b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4284b087196eSStefano Zampini PetscScalar val; 4285b087196eSStefano Zampini PetscInt idx; 4286b087196eSStefano Zampini 4287b087196eSStefano Zampini idx = idxs_V[i]; 4288b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4289b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4290b087196eSStefano Zampini } 4291b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4292b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4293bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4294bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4295bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4296bf3a8328SStefano Zampini } 4297ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4298ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4299ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4300ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4301b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4302ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4303bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4304b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4305bf3a8328SStefano Zampini } 4306ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4307ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4308ac632422SStefano Zampini } 4309b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 431088c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4311b96c3477SStefano Zampini } 4312c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4313c9db6a07SStefano Zampini if (sub_schurs->change) { 4314c9db6a07SStefano Zampini PetscInt i; 4315c9db6a07SStefano Zampini 4316c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4317c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4318c9db6a07SStefano Zampini } 4319c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4320c9db6a07SStefano Zampini } 4321b96c3477SStefano Zampini } 432216909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 432316909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 432416909a7fSStefano Zampini } else { 4325906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 432616909a7fSStefano Zampini } 43271dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 43281dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43291dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43301dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->change_interior,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43311dd7afcfSStefano Zampini } 433272b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 433372b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 433472b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 433572b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 433672b8c272SStefano Zampini } 43371dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 43381dd7afcfSStefano Zampini PetscBool needglobal = PETSC_FALSE; 43391dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43401dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43411dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&needglobal,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43421dd7afcfSStefano Zampini } 43431dd7afcfSStefano Zampini if (!needglobal && pcbddc->user_ChangeOfBasisMatrix) { 4344b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4345b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4346906d46d4SStefano Zampini } else { 43471dd7afcfSStefano Zampini Mat benign_global = NULL; 43481dd7afcfSStefano Zampini if (needglobal) { 43491dd7afcfSStefano Zampini Mat tmat; 43501dd7afcfSStefano Zampini 43511dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 43521dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 43531dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 43541dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43551dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43561dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 43571dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43581dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43591dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 43601dd7afcfSStefano Zampini if (pcbddc->benign_change) { 43611dd7afcfSStefano Zampini Mat M; 43621dd7afcfSStefano Zampini 43631dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 43641dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 43651dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 43661dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4367906d46d4SStefano Zampini } else { 43681dd7afcfSStefano Zampini Mat eye; 43691dd7afcfSStefano Zampini PetscScalar *array; 43701dd7afcfSStefano Zampini 43711dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43721dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 43731dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 43741dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4375906d46d4SStefano Zampini } 43761dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43771dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43781dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43791dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 43801dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 43811dd7afcfSStefano Zampini } 43821dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 43831dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 43841dd7afcfSStefano Zampini } 43851dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 43861dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 43871dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 43881dd7afcfSStefano Zampini } else if (needglobal) { 43891dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 43901dd7afcfSStefano Zampini } 43911dd7afcfSStefano Zampini } 439216909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 439316909a7fSStefano Zampini IS is_global; 439416909a7fSStefano Zampini const PetscInt *gidxs; 439516909a7fSStefano Zampini 439616909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 439716909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 439816909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 439916909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 440016909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 440116909a7fSStefano Zampini } 44021dd7afcfSStefano Zampini } 44031dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 44041dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4405b9b85e73SStefano Zampini } 4406a717540cSStefano Zampini 440772b8c272SStefano Zampini if (!pcbddc->fake_change) { 44084f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 44094f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 44104f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 44114f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4412019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4413019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4414019a44ceSStefano Zampini pcbddc->local_primal_size++; 4415019a44ceSStefano Zampini } 4416019a44ceSStefano Zampini 4417019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4418727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4419727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 44209f47a83aSStefano 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); 4421c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 44220e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 44239f47a83aSStefano 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); 4424727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4425727cdba6SStefano Zampini } 44260e6343abSStefano Zampini } 4427727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4428b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 442972b8c272SStefano Zampini } 443072b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4431727cdba6SStefano Zampini 4432a717540cSStefano Zampini /* flush dbg viewer */ 4433b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4434b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4435b8ffe317SStefano Zampini } 4436a717540cSStefano Zampini 4437e310c8b4SStefano Zampini /* free workspace */ 4438a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 44394641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 444008122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 44419162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 44429162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 444308122e43SStefano Zampini } else { 44449162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 44459162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 44469162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 444708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 444808122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 44499162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 44509162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 445108122e43SStefano Zampini } 4452674ae819SStefano Zampini PetscFunctionReturn(0); 4453674ae819SStefano Zampini } 4454674ae819SStefano Zampini 4455674ae819SStefano Zampini #undef __FUNCT__ 4456674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4457674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4458674ae819SStefano Zampini { 4459674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4460674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4461674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 446214f95afaSStefano Zampini PetscInt ierr,i,N; 4463674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4464674ae819SStefano Zampini 4465674ae819SStefano Zampini PetscFunctionBegin; 44668e61c736SStefano Zampini /* Reset previously computed graph */ 44678e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4468674ae819SStefano Zampini /* Init local Graph struct */ 44697fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 44703bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4471674ae819SStefano Zampini 4472575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 44735099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 44745099eff2SStefano 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); 4475575ad6abSStefano Zampini } 44769577ea80SStefano Zampini 4477674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4478d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 44794d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 44804d379d7bSStefano Zampini PetscInt nvtxs; 4481e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4482674ae819SStefano Zampini 44832fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 44842fffb893SStefano Zampini if (flg_row) { 44854d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4486b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 44872fffb893SStefano Zampini } 44882fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4489674ae819SStefano Zampini } 44909b28b941SStefano Zampini if (pcbddc->dbg_flag) { 44919b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4492674ae819SStefano Zampini } 4493674ae819SStefano Zampini 449463602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 449563602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 449663602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 449795ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 449863602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4499e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 450063602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4501674ae819SStefano Zampini } 450263602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 450363602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 450463602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4505674ae819SStefano Zampini } 450663602bcaSStefano Zampini } else { 450763602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 450863602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4509854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 451063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 451163602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 451263602bcaSStefano Zampini } 451363602bcaSStefano Zampini } 4514674ae819SStefano Zampini } 4515674ae819SStefano Zampini 4516674ae819SStefano Zampini /* Setup of Graph */ 4517785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4518e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4519785d1243SStefano Zampini } 4520785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4521e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4522785d1243SStefano Zampini } 452330368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 452430368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 452530368db7SStefano Zampini } 452614f95afaSStefano 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); 4527674ae819SStefano Zampini 45284f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 45294f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 45304f1b2e48SStefano Zampini PetscInt *local_subs; 45314f1b2e48SStefano Zampini 45324f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 45334f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 45344f1b2e48SStefano Zampini const PetscInt *idxs; 45354f1b2e48SStefano Zampini PetscInt nl,j; 45364f1b2e48SStefano Zampini 45374f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 45384f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45394f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 45404f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 45414f1b2e48SStefano Zampini } 45424f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45434f1b2e48SStefano Zampini } 45444f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 45454f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 45464f1b2e48SStefano Zampini } 45474f1b2e48SStefano Zampini 4548674ae819SStefano Zampini /* Graph's connected components analysis */ 4549674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4550674ae819SStefano Zampini 4551674ae819SStefano Zampini /* print some info to stdout */ 4552674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4553302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4554674ae819SStefano Zampini } 4555674ae819SStefano Zampini PetscFunctionReturn(0); 4556674ae819SStefano Zampini } 4557674ae819SStefano Zampini 4558dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4559674ae819SStefano Zampini #undef __FUNCT__ 4560674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4561dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4562674ae819SStefano Zampini { 4563dc456d91SStefano Zampini PetscSF sf; 4564dc456d91SStefano Zampini PetscLayout map; 4565dc456d91SStefano Zampini const PetscInt *idxs; 4566dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4567dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4568dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4569dc456d91SStefano Zampini PetscMPIInt commsize; 4570674ae819SStefano Zampini PetscBool first_found; 4571674ae819SStefano Zampini PetscErrorCode ierr; 4572674ae819SStefano Zampini 4573674ae819SStefano Zampini PetscFunctionBegin; 4574dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4575dc456d91SStefano Zampini if (subset_mult) { 4576dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4577dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4578dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4579674ae819SStefano Zampini } 4580dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4581dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4582dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4583dc456d91SStefano Zampini for (i=0;i<n;i++) { 4584dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4585dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4586674ae819SStefano Zampini } 4587dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4588b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4589dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4590dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4591dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4592dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4593dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4594dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4595dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4596dc456d91SStefano Zampini 4597dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4598dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4599dc456d91SStefano Zampini if (subset_mult) { 4600dc456d91SStefano Zampini const PetscInt* idxs_mult; 4601dc456d91SStefano Zampini 4602dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4603dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4604dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4605674ae819SStefano Zampini } else { 4606dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4607674ae819SStefano Zampini } 4608dc456d91SStefano Zampini /* local size of new subset */ 4609dc456d91SStefano Zampini n_n = 0; 4610dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4611dc456d91SStefano Zampini 4612dc456d91SStefano Zampini /* global indexes in layout */ 4613dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4614dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4615dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4616dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4617dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4618dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4619dc456d91SStefano Zampini 4620dc456d91SStefano Zampini /* reduce from leaves to roots */ 4621dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 462264a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 462364a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4624dc456d91SStefano Zampini 4625dc456d91SStefano Zampini /* count indexes in local part of layout */ 4626674ae819SStefano Zampini nlocals = 0; 4627674ae819SStefano Zampini first_index = -1; 4628674ae819SStefano Zampini first_found = PETSC_FALSE; 4629dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4630dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4631674ae819SStefano Zampini first_found = PETSC_TRUE; 4632674ae819SStefano Zampini first_index = i; 4633674ae819SStefano Zampini } 4634dc456d91SStefano Zampini nlocals += root_data[i]; 4635674ae819SStefano Zampini } 4636dc456d91SStefano Zampini 4637dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 46385fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4639dc456d91SStefano Zampini start = 0; 464064a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46415fa240b1SStefano Zampini #else 464264a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46435fa240b1SStefano Zampini start = start-nlocals; 46445fa240b1SStefano Zampini #endif 46455fa240b1SStefano Zampini 4646dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4647dc456d91SStefano Zampini *N_n = start + nlocals; 4648dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4649dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4650674ae819SStefano Zampini } 46515fa240b1SStefano Zampini 46525fa240b1SStefano Zampini /* adapt root data with cumulative */ 4653674ae819SStefano Zampini if (first_found) { 4654dc456d91SStefano Zampini PetscInt old_index; 4655dc456d91SStefano Zampini 4656dc456d91SStefano Zampini root_data[first_index] += start; 4657674ae819SStefano Zampini old_index = first_index; 4658dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4659dc456d91SStefano Zampini if (root_data[i]) { 4660dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4661674ae819SStefano Zampini old_index = i; 4662674ae819SStefano Zampini } 4663674ae819SStefano Zampini } 4664674ae819SStefano Zampini } 4665dc456d91SStefano Zampini 4666dc456d91SStefano Zampini /* from roots to leaves */ 4667dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4668dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4669dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4670dc456d91SStefano Zampini 4671dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4672dc456d91SStefano Zampini if (subset_mult) { 4673dc456d91SStefano Zampini const PetscInt* idxs_mult; 4674dc456d91SStefano Zampini PetscInt cum; 4675dc456d91SStefano Zampini 4676dc456d91SStefano Zampini cum = 0; 4677dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4678dc456d91SStefano Zampini for (i=0;i<n;i++) { 4679dc456d91SStefano Zampini PetscInt j; 4680dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4681674ae819SStefano Zampini } 4682dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4683674ae819SStefano Zampini } else { 4684dc456d91SStefano Zampini for (i=0;i<n;i++) { 4685dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4686674ae819SStefano Zampini } 4687674ae819SStefano Zampini } 4688dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4689dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4690674ae819SStefano Zampini PetscFunctionReturn(0); 4691674ae819SStefano Zampini } 46929a7d3425SStefano Zampini 46939a7d3425SStefano Zampini #undef __FUNCT__ 46949a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 46959a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 46969a7d3425SStefano Zampini { 46979a7d3425SStefano Zampini PetscInt i,j; 46989a7d3425SStefano Zampini PetscScalar *alphas; 46999a7d3425SStefano Zampini PetscErrorCode ierr; 47009a7d3425SStefano Zampini 47019a7d3425SStefano Zampini PetscFunctionBegin; 47029a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4703785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 47049a7d3425SStefano Zampini for (i=0;i<n;i++) { 47059a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 47069a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 47079a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 47089a7d3425SStefano Zampini } 47099a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 47109a7d3425SStefano Zampini PetscFunctionReturn(0); 47119a7d3425SStefano Zampini } 47129a7d3425SStefano Zampini 4713e7931f94SStefano Zampini #undef __FUNCT__ 471470cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4715b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4716e7931f94SStefano Zampini { 471752e5ac9dSStefano Zampini IS ranks_send_to; 4718e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4719e7931f94SStefano Zampini PetscMPIInt size,rank,color; 472052e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 472152e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 47223837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 47232b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4724e7931f94SStefano Zampini PetscSubcomm subcomm; 472552e5ac9dSStefano Zampini PetscErrorCode ierr; 4726a57a6d2fSStefano Zampini 4727e7931f94SStefano Zampini PetscFunctionBegin; 4728c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 4729c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4730c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4731e7931f94SStefano Zampini 4732e7931f94SStefano Zampini /* Get info on mapping */ 47333bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 47343bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4735e7931f94SStefano Zampini 4736e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4737785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4738e7931f94SStefano Zampini xadj[0] = 0; 4739e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4740785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4741785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4742e7931f94SStefano Zampini 47432b510759SStefano Zampini if (threshold) { 4744d023bfaeSStefano Zampini PetscInt xadj_count = 0; 47452b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4746d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4747d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4748d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4749d023bfaeSStefano Zampini xadj_count++; 4750e7931f94SStefano Zampini } 4751e7931f94SStefano Zampini } 4752d023bfaeSStefano Zampini xadj[1] = xadj_count; 4753c8587f34SStefano Zampini } else { 4754e7931f94SStefano Zampini if (xadj[1]) { 4755e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4756e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4757c8587f34SStefano Zampini } 4758e7931f94SStefano Zampini } 47593bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4760e7931f94SStefano Zampini if (use_square) { 4761e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4762e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4763e7931f94SStefano Zampini } 4764e7931f94SStefano Zampini } 4765e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4766e7931f94SStefano Zampini 47673837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4768e7931f94SStefano Zampini 4769e7931f94SStefano Zampini /* 4770e7931f94SStefano Zampini Restrict work on active processes only. 4771e7931f94SStefano Zampini */ 4772e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4773e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4774e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 47752b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4776d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4777e7931f94SStefano Zampini if (color) { 4778e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4779e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4780e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4781c8587f34SStefano Zampini } else { 478252e5ac9dSStefano Zampini Mat subdomain_adj; 478352e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 478452e5ac9dSStefano Zampini MatPartitioning partitioner; 478552e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 478652e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4787b0c7d250SStefano Zampini PetscBool aggregate; 4788b0c7d250SStefano Zampini 4789306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4790785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4791e7931f94SStefano Zampini prank = rank; 4792306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 47938002ef2cSStefano Zampini /* 4794e7931f94SStefano Zampini for (i=0;i<size;i++) { 4795e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4796c8587f34SStefano Zampini } 47978002ef2cSStefano Zampini */ 4798e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4799e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4800c8587f34SStefano Zampini } 4801e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4802b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4803b0c7d250SStefano Zampini if (aggregate) { 4804b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4805b0c7d250SStefano Zampini PetscMPIInt nrank; 4806b0c7d250SStefano Zampini PetscScalar *vals; 4807b0c7d250SStefano Zampini 4808b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4809b0c7d250SStefano Zampini lrows = 0; 4810b0c7d250SStefano Zampini if (nrank<redprocs) { 4811b0c7d250SStefano Zampini lrows = size/redprocs; 4812b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4813b0c7d250SStefano Zampini } 48145fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4815b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4816b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4817b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4818b0c7d250SStefano Zampini row = nrank; 4819b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4820b0c7d250SStefano Zampini cols = adjncy; 4821b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4822b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4823b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4824b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4825b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 482652e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 482752e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 482852e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4829b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4830b0c7d250SStefano Zampini } else { 4831306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4832b0c7d250SStefano Zampini } 483322b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4834e7931f94SStefano Zampini 4835e7931f94SStefano Zampini /* Partition */ 4836306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4837e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4838e7931f94SStefano Zampini if (use_vwgt) { 48393837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4840e7931f94SStefano Zampini v_wgt[0] = local_size; 4841e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4842c8587f34SStefano Zampini } 484328143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 484428143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4845e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4846e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 484722b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4848e7931f94SStefano Zampini 484952e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 485052e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 485152e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 485252e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4853b0c7d250SStefano Zampini if (!redprocs) { 4854b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 485528143c3dSStefano Zampini } else { 4856b0c7d250SStefano Zampini PetscInt idxs[1]; 4857b0c7d250SStefano Zampini PetscMPIInt tag; 4858b0c7d250SStefano Zampini MPI_Request *reqs; 4859b0c7d250SStefano Zampini 4860b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4861b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4862b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4863b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 486428143c3dSStefano Zampini } 4865b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4866b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4867b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4868b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4869e7931f94SStefano Zampini } 487052e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4871e7931f94SStefano Zampini /* clean up */ 4872e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 487352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4874e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4875e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4876e7931f94SStefano Zampini } 4877e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4878e7931f94SStefano Zampini 4879e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4880e7931f94SStefano Zampini i = 1; 4881e7931f94SStefano Zampini if (color) i=0; 4882e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4883e7931f94SStefano Zampini /* get back IS */ 4884e7931f94SStefano Zampini *is_sends = ranks_send_to; 4885e7931f94SStefano Zampini PetscFunctionReturn(0); 4886e7931f94SStefano Zampini } 4887e7931f94SStefano Zampini 4888e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4889e7931f94SStefano Zampini 4890e7931f94SStefano Zampini #undef __FUNCT__ 4891e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 489253a05cb3SStefano Zampini PetscErrorCode MatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, MatReuse reuse, Mat *mat_n, PetscInt nis, IS isarray[]) 4893e7931f94SStefano Zampini { 489470cf5478SStefano Zampini Mat local_mat; 4895e7931f94SStefano Zampini IS is_sends_internal; 48969d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 489728143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 48989d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4899e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4900e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4901e7931f94SStefano Zampini const PetscInt* is_indices; 4902e7931f94SStefano Zampini MatType new_local_type; 4903e7931f94SStefano Zampini /* buffers */ 4904e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 490528143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 49069d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4907e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4908e7931f94SStefano Zampini /* MPI */ 490928143c3dSStefano Zampini MPI_Comm comm,comm_n; 491028143c3dSStefano Zampini PetscSubcomm subcomm; 4911e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 491228143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 491328143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 491428143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 491528143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 491628143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4917e7931f94SStefano Zampini PetscErrorCode ierr; 4918e7931f94SStefano Zampini 4919e7931f94SStefano Zampini PetscFunctionBegin; 492028143c3dSStefano Zampini /* TODO: add missing checks */ 492128143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 492228143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 492328143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 492428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4925e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 492628143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4927e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4928e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4929e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4930e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4931e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 493228143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 493370cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 493470cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 493528143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 493670cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 493770cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 493870cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 493970cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 494070cf5478SStefano Zampini } 4941e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4942e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4943e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4944e7931f94SStefano Zampini if (!is_sends) { 494528143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4946b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4947c8587f34SStefano Zampini } else { 4948e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4949e7931f94SStefano Zampini is_sends_internal = is_sends; 4950c8587f34SStefano Zampini } 4951e7931f94SStefano Zampini 4952e7931f94SStefano Zampini /* get comm */ 4953a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4954e7931f94SStefano Zampini 4955e7931f94SStefano Zampini /* compute number of sends */ 4956e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4957e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4958e7931f94SStefano Zampini 4959e7931f94SStefano Zampini /* compute number of receives */ 4960e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4961785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4962e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4963e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4964e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4965e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4966e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4967e7931f94SStefano Zampini 496828143c3dSStefano Zampini /* restrict comm if requested */ 496928143c3dSStefano Zampini subcomm = 0; 497028143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 497128143c3dSStefano Zampini if (restrict_comm) { 4972779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4973779c1cceSStefano Zampini 497428143c3dSStefano Zampini color = 0; 497553a05cb3SStefano Zampini if (restrict_full) { 497653a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 497753a05cb3SStefano Zampini } else { 497853a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 497953a05cb3SStefano Zampini } 4980b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 498128143c3dSStefano Zampini subcommsize = commsize - subcommsize; 498228143c3dSStefano Zampini /* check if reuse has been requested */ 498328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 498428143c3dSStefano Zampini if (*mat_n) { 498528143c3dSStefano Zampini PetscMPIInt subcommsize2; 498628143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 498728143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 498828143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 498928143c3dSStefano Zampini } else { 499028143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 499128143c3dSStefano Zampini } 499228143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4993779c1cceSStefano Zampini PetscMPIInt rank; 4994779c1cceSStefano Zampini 4995779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 499628143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 499728143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 499828143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4999306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 500028143c3dSStefano Zampini } 500128143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 500228143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 500328143c3dSStefano Zampini } else { 500428143c3dSStefano Zampini comm_n = comm; 500528143c3dSStefano Zampini } 500628143c3dSStefano Zampini 5007e7931f94SStefano Zampini /* prepare send/receive buffers */ 5008785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5009e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5010785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5011e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 501228143c3dSStefano Zampini if (nis) { 5013854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 501428143c3dSStefano Zampini } 5015e7931f94SStefano Zampini 501628143c3dSStefano Zampini /* Get data from local matrices */ 50176c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5018e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5019e7931f94SStefano Zampini /* 5020e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5021e7931f94SStefano Zampini send_buffer_idxs should contain: 5022e7931f94SStefano Zampini - MatType_PRIVATE type 5023e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5024e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5025e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5026e7931f94SStefano Zampini */ 50276c4ed002SBarry Smith else { 5028e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 50293bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5030854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5031e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5032e7931f94SStefano Zampini send_buffer_idxs[1] = i; 50333bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5034e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 50353bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5036e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5037e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5038e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5039e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5040c8587f34SStefano Zampini } 5041c8587f34SStefano Zampini } 5042e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 504328143c3dSStefano Zampini /* additional is (if any) */ 504428143c3dSStefano Zampini if (nis) { 504528143c3dSStefano Zampini PetscMPIInt psum; 504628143c3dSStefano Zampini PetscInt j; 504728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 504828143c3dSStefano Zampini PetscInt plen; 504928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 505028143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 505128143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 505228143c3dSStefano Zampini } 5053854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 505428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 505528143c3dSStefano Zampini PetscInt plen; 505628143c3dSStefano Zampini const PetscInt *is_array_idxs; 505728143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 505828143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 505928143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 506028143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 506128143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 506228143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 506328143c3dSStefano Zampini } 506428143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 506528143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 506628143c3dSStefano Zampini } 506728143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 506828143c3dSStefano Zampini } 506928143c3dSStefano Zampini 5070e7931f94SStefano Zampini buf_size_idxs = 0; 5071e7931f94SStefano Zampini buf_size_vals = 0; 507228143c3dSStefano Zampini buf_size_idxs_is = 0; 5073e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5074e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5075e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 507628143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5077e7931f94SStefano Zampini } 5078785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5079785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 508095ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5081e7931f94SStefano Zampini 5082e7931f94SStefano Zampini /* get new tags for clean communications */ 5083e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5084e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 508528143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5086e7931f94SStefano Zampini 5087e7931f94SStefano Zampini /* allocate for requests */ 5088785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5089785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 509095ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5091785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5092785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 509395ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5094e7931f94SStefano Zampini 5095e7931f94SStefano Zampini /* communications */ 5096e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5097e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 509828143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5099e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5100e7931f94SStefano Zampini source_dest = onodes[i]; 5101e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5102e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5103e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5104e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 510528143c3dSStefano Zampini if (nis) { 510628143c3dSStefano 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); 510728143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 510828143c3dSStefano Zampini } 5109e7931f94SStefano Zampini } 5110e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5111e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5112e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5113e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 511428143c3dSStefano Zampini if (nis) { 511528143c3dSStefano 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); 511628143c3dSStefano Zampini } 5117e7931f94SStefano Zampini } 5118e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5119e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5120e7931f94SStefano Zampini 5121e7931f94SStefano Zampini /* assemble new l2g map */ 5122e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5123e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51249d30be91SStefano Zampini new_local_rows = 0; 5125e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51269d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5127e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5128e7931f94SStefano Zampini } 51299d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5130e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51319d30be91SStefano Zampini new_local_rows = 0; 5132e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51339d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 51349d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5135e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5136e7931f94SStefano Zampini } 51379d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 51389d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5139e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5140e7931f94SStefano Zampini 5141e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5142e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5143e7931f94SStefano 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) */ 5144e7931f94SStefano Zampini if (n_recvs) { 514528143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5146e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5147e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5148e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5149e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5150e7931f94SStefano Zampini break; 5151e7931f94SStefano Zampini } 5152e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5153e7931f94SStefano Zampini } 5154e7931f94SStefano Zampini switch (new_local_type_private) { 515528143c3dSStefano Zampini case MATDENSE_PRIVATE: 515628143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5157e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5158e7931f94SStefano Zampini bs = 1; 515928143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 516028143c3dSStefano Zampini new_local_type = MATSEQDENSE; 516128143c3dSStefano Zampini bs = 1; 516228143c3dSStefano Zampini } 5163e7931f94SStefano Zampini break; 5164e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5165e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5166e7931f94SStefano Zampini bs = 1; 5167e7931f94SStefano Zampini break; 5168e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5169e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5170e7931f94SStefano Zampini break; 5171e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5172e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5173e7931f94SStefano Zampini break; 5174e7931f94SStefano Zampini default: 51759d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5176e7931f94SStefano Zampini break; 5177e7931f94SStefano Zampini } 517828143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 517928143c3dSStefano Zampini new_local_type = MATSEQDENSE; 518028143c3dSStefano Zampini bs = 1; 5181e7931f94SStefano Zampini } 5182e7931f94SStefano Zampini 518370cf5478SStefano Zampini /* create MATIS object if needed */ 518470cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5185e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5186e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 518770cf5478SStefano Zampini } else { 518870cf5478SStefano Zampini /* it also destroys the local matrices */ 518970cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 519070cf5478SStefano Zampini } 519170cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5192e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 51939d30be91SStefano Zampini 51949d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 51959d30be91SStefano Zampini 51969d30be91SStefano Zampini /* Global to local map of received indices */ 51979d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 51989d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 51999d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 52009d30be91SStefano Zampini 52019d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 52029d30be91SStefano Zampini buf_size_idxs = 0; 52039d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 52049d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 52059d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 52069d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 52079d30be91SStefano Zampini } 52089d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 52099d30be91SStefano Zampini 52109d30be91SStefano Zampini /* set preallocation */ 52119d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 52129d30be91SStefano Zampini if (!newisdense) { 52139d30be91SStefano Zampini PetscInt *new_local_nnz=0; 52149d30be91SStefano Zampini 52159d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 52169d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 52179d30be91SStefano Zampini if (n_recvs) { 52189d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 52199d30be91SStefano Zampini } 52209d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 52219d30be91SStefano Zampini PetscInt j; 52229d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 52239d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 52249d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 52259d30be91SStefano Zampini } 52269d30be91SStefano Zampini } else { 52279d30be91SStefano Zampini /* TODO */ 52289d30be91SStefano Zampini } 52299d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 52309d30be91SStefano Zampini } 52319d30be91SStefano Zampini if (new_local_nnz) { 52329d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 52339d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 52349d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 52359d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52369d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 52379d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52389d30be91SStefano Zampini } else { 52399d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52409d30be91SStefano Zampini } 52419d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 52429d30be91SStefano Zampini } else { 52439d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52449d30be91SStefano Zampini } 5245e7931f94SStefano Zampini 5246e7931f94SStefano Zampini /* set values */ 5247e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 52489d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5249e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5250e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5251e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 52529d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5253e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5254e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5255e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 525628143c3dSStefano Zampini } else { 525728143c3dSStefano Zampini /* TODO */ 5258e7931f94SStefano Zampini } 5259e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5260e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5261e7931f94SStefano Zampini } 5262e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5263e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 526470cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 526570cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 52669d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 52679d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5268e7931f94SStefano Zampini 5269dfd14d43SStefano Zampini #if 0 527028143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5271e7931f94SStefano Zampini Vec lvec,rvec; 5272e7931f94SStefano Zampini PetscReal infty_error; 5273e7931f94SStefano Zampini 52742a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5275e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5276e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5277e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 527870cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5279e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5280e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5281e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5282e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5283e7931f94SStefano Zampini } 528428143c3dSStefano Zampini #endif 5285e7931f94SStefano Zampini 528628143c3dSStefano Zampini /* assemble new additional is (if any) */ 528728143c3dSStefano Zampini if (nis) { 528828143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 528928143c3dSStefano Zampini 529028143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5291854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 529228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 529328143c3dSStefano Zampini psum = 0; 529428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 529528143c3dSStefano Zampini for (j=0;j<nis;j++) { 529628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 529728143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 529828143c3dSStefano Zampini psum += plen; 529928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 530028143c3dSStefano Zampini } 530128143c3dSStefano Zampini } 5302854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5303854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 530428143c3dSStefano Zampini for (i=1;i<nis;i++) { 530528143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 530628143c3dSStefano Zampini } 530728143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 530828143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 530928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 531028143c3dSStefano Zampini for (j=0;j<nis;j++) { 531128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 531228143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 531328143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 531428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 531528143c3dSStefano Zampini } 531628143c3dSStefano Zampini } 531728143c3dSStefano Zampini for (i=0;i<nis;i++) { 531828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 531928143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 532028143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 532128143c3dSStefano Zampini } 532228143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 532328143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 532428143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 532528143c3dSStefano Zampini } 5326e7931f94SStefano Zampini /* free workspace */ 532728143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5328e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5329e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5330e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5331e7931f94SStefano Zampini if (isdense) { 5332e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5333e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5334e7931f94SStefano Zampini } else { 5335e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5336e7931f94SStefano Zampini } 533728143c3dSStefano Zampini if (nis) { 533828143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 533928143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 534028143c3dSStefano Zampini } 5341e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5342e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 534328143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5344e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5345e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 534628143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5347e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5348e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5349e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5350e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5351e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 535228143c3dSStefano Zampini if (nis) { 535328143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 535428143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 535528143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 535628143c3dSStefano Zampini } 535728143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 535828143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 535928143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 536028143c3dSStefano Zampini for (i=0;i<nis;i++) { 536128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 536228143c3dSStefano Zampini } 536353a05cb3SStefano Zampini *mat_n = NULL; 536428143c3dSStefano Zampini } 5365e7931f94SStefano Zampini PetscFunctionReturn(0); 5366e7931f94SStefano Zampini } 5367a57a6d2fSStefano Zampini 536812edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5369af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 537012edc857SStefano Zampini 5371c8587f34SStefano Zampini #undef __FUNCT__ 5372c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5373c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5374c8587f34SStefano Zampini { 5375c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5376c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 537720a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 53789881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 537920a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 53806e683305SStefano Zampini IS coarse_is,*isarray; 53816e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 538230368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5383f9eb5b7dSStefano Zampini PC pc_temp; 5384c8587f34SStefano Zampini PCType coarse_pc_type; 5385c8587f34SStefano Zampini KSPType coarse_ksp_type; 5386f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 53874f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 53886e683305SStefano Zampini Mat t_coarse_mat_is; 53896e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 53906e683305SStefano Zampini PetscMPIInt all_procs; 539174e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 539268457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 539322bc73bbSStefano Zampini PetscScalar *array; 53949881197aSStefano Zampini PetscErrorCode ierr; 5395fdc09c96SStefano Zampini 5396c8587f34SStefano Zampini PetscFunctionBegin; 5397c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 539868457ee5SStefano 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 */ 5399fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 54005a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5401fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5402f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5403f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5404f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5405fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 540651bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 540751bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5408dc4bcba2SStefano Zampini PC pc; 5409dc4bcba2SStefano Zampini PetscBool isbddc; 5410dc4bcba2SStefano Zampini 5411dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5412dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5413dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5414dc4bcba2SStefano Zampini if (isbddc) { 541563c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 541663c961adSStefano Zampini } else { 5417727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 541863c961adSStefano Zampini } 5419fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5420fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5421fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5422f4ddd8eeSStefano Zampini } 5423fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5424fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5425f4ddd8eeSStefano Zampini } 542670cf5478SStefano Zampini /* reset any subassembling information */ 542770cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 54286e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54296e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5430fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5431f4ddd8eeSStefano Zampini } 5432c8587f34SStefano Zampini 54336e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 54342b510759SStefano Zampini im_active = !!(pcis->n); 5435b2566f29SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 54366e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 54376e683305SStefano Zampini void_procs = all_procs-active_procs; 54386e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 543974e2c79eSStefano Zampini redist = PETSC_FALSE; 544022bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 54416e683305SStefano Zampini csin_ml = PETSC_TRUE; 54426e683305SStefano Zampini ncoarse_ml = void_procs; 5443779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5444779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 54456e683305SStefano Zampini csin_ds = PETSC_TRUE; 544618a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 544718a45a71SStefano Zampini redist = PETSC_TRUE; 544818a45a71SStefano Zampini } else { 54496e683305SStefano Zampini csin_ds = PETSC_TRUE; 5450779c1cceSStefano Zampini ncoarse_ds = active_procs; 5451779c1cceSStefano Zampini redist = PETSC_TRUE; 545218a45a71SStefano Zampini } 54536e683305SStefano Zampini } else { 54546e683305SStefano Zampini csin_ml = PETSC_FALSE; 54556e683305SStefano Zampini ncoarse_ml = all_procs; 54566e683305SStefano Zampini if (void_procs) { 54576e683305SStefano Zampini csin_ds = PETSC_TRUE; 54586e683305SStefano Zampini ncoarse_ds = void_procs; 54596e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 54606e683305SStefano Zampini } else { 5461779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 546274e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 546374e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 546474e2c79eSStefano Zampini redist = PETSC_TRUE; 546574e2c79eSStefano Zampini } else { 54666e683305SStefano Zampini csin_ds = PETSC_FALSE; 54676e683305SStefano Zampini ncoarse_ds = all_procs; 54686e683305SStefano Zampini } 54696e683305SStefano Zampini } 547074e2c79eSStefano Zampini } 54716e683305SStefano Zampini 54726e683305SStefano Zampini /* 54736e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 54746e683305SStefano Zampini - we have not exceeded the number of levels requested 54756e683305SStefano Zampini - we can actually subassemble the active processes 54766e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 54776e683305SStefano Zampini */ 54786e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 54796e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 54806e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 54816e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 54826e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5483f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 54842b510759SStefano Zampini } else { 5485f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5486c8587f34SStefano Zampini } 5487c8587f34SStefano Zampini } 54886e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 54896e683305SStefano Zampini if (multilevel_allowed) { 54906e683305SStefano Zampini ncoarse = ncoarse_ml; 54916e683305SStefano Zampini csin = csin_ml; 549258da7f69SStefano Zampini redist = PETSC_FALSE; 54936e683305SStefano Zampini } else { 54946e683305SStefano Zampini ncoarse = ncoarse_ds; 54956e683305SStefano Zampini csin = csin_ds; 54966e683305SStefano Zampini } 5497e7931f94SStefano Zampini 5498abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5499abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5500abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5501abbbba34SStefano Zampini 5502abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 550322bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 550422bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 550522bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 550622bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5507e176bc59SStefano 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); 55086e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 55096e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55106e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5511abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5512abbbba34SStefano Zampini 55136e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 551430368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 55156e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 55166e683305SStefano Zampini const PetscInt *idxs; 55176e683305SStefano Zampini ISLocalToGlobalMapping tmap; 55186e683305SStefano Zampini 55196e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 55200be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 55216e683305SStefano Zampini /* allocate space for temporary storage */ 5522854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5523854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 55246e683305SStefano Zampini /* allocate for IS array */ 55256e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 55266e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 552730368db7SStefano Zampini nisvert = 0; 552830368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 552930368db7SStefano Zampini nisvert = 1; 553030368db7SStefano Zampini } 553130368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5532854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 55336e683305SStefano Zampini /* dofs splitting */ 55346e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 55356e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 55366e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 55376e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55386e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55396e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55406e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 554130368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 55426e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 55436e683305SStefano Zampini } 55446e683305SStefano Zampini /* neumann boundaries */ 55456e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 55466e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 55476e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 55486e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55496e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55506e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55516e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 555230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 55536e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 55546e683305SStefano Zampini } 555530368db7SStefano Zampini /* primal vertices (benign) */ 555630368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 555730368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 555830368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 555930368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 556030368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 556130368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 556230368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 556330368db7SStefano Zampini } 55646e683305SStefano Zampini /* free memory */ 55656e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 55666e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 55676e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 55686e683305SStefano Zampini } else { 55696e683305SStefano Zampini nis = 0; 55706e683305SStefano Zampini nisdofs = 0; 55716e683305SStefano Zampini nisneu = 0; 557230368db7SStefano Zampini nisvert = 0; 55736e683305SStefano Zampini isarray = NULL; 55746e683305SStefano Zampini } 55756e683305SStefano Zampini /* destroy no longer needed map */ 55766e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 55776e683305SStefano Zampini 55786e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 55796e683305SStefano Zampini coarse_mat_is = NULL; 55806e683305SStefano Zampini if (csin) { 55816e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 558274e2c79eSStefano Zampini if (redist) { 558374e2c79eSStefano Zampini PetscMPIInt rank; 5584779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 558574e2c79eSStefano Zampini 558674e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 558758da7f69SStefano Zampini spc = active_procs/ncoarse; 558858da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5589779c1cceSStefano Zampini if (im_active) { 5590779c1cceSStefano Zampini destsize = 1; 559174e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 559274e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 559374e2c79eSStefano Zampini } else { 559474e2c79eSStefano Zampini dest[0] = rank/(spc+1); 559574e2c79eSStefano Zampini } 559674e2c79eSStefano Zampini } else { 5597779c1cceSStefano Zampini destsize = 0; 55986e683305SStefano Zampini } 5599779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5600779c1cceSStefano Zampini } else if (csin_type_simple) { 56016e683305SStefano Zampini PetscMPIInt rank; 56026e683305SStefano Zampini PetscInt issize,isidx; 5603779c1cceSStefano Zampini 56046e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 56056e683305SStefano Zampini if (im_active) { 56066e683305SStefano Zampini issize = 1; 56076e683305SStefano Zampini isidx = (PetscInt)rank; 56086e683305SStefano Zampini } else { 56096e683305SStefano Zampini issize = 0; 56106e683305SStefano Zampini isidx = -1; 56116e683305SStefano Zampini } 56126e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5613779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5614b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 56156e683305SStefano Zampini } 5616779c1cceSStefano Zampini 5617779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5618779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5619779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5620779c1cceSStefano Zampini PetscInt *coarse_candidates; 5621779c1cceSStefano Zampini const PetscInt* tisindices; 5622779c1cceSStefano Zampini 5623779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5624779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5625779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5626779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5627779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5628779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5629779c1cceSStefano Zampini } 5630779c1cceSStefano Zampini } 5631779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5632779c1cceSStefano Zampini 5633779c1cceSStefano Zampini 56346e683305SStefano Zampini if (pcbddc->dbg_flag) { 56356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 56376e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56386e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5639779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 56406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 56416e683305SStefano Zampini } 56426e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 56436e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56446e683305SStefano Zampini } 56456e683305SStefano Zampini /* shift the pattern on coarse candidates */ 56466e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 56476e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5648854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 56496e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 56506e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 56516e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 56526e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 56536e683305SStefano Zampini } 56546e683305SStefano Zampini if (pcbddc->dbg_flag) { 56556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56566e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 56576e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56586e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56596e683305SStefano Zampini } 5660779c1cceSStefano Zampini } 56616e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 566253a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 566353a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_FALSE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 566453a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 566553a05cb3SStefano Zampini ierr = MatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling_init,0,PETSC_TRUE,PETSC_TRUE,MAT_INITIAL_MATRIX,&coarse_mat_is,nis,isarray);CHKERRQ(ierr); 566653a05cb3SStefano Zampini } 56676e683305SStefano Zampini } else { 56686e683305SStefano Zampini if (pcbddc->dbg_flag) { 56696e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 56716e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56726e683305SStefano Zampini } 56736e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 56746e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 56756e683305SStefano Zampini } 56766e683305SStefano Zampini 56776e683305SStefano Zampini /* create local to global scatters for coarse problem */ 567868457ee5SStefano Zampini if (compute_vecs) { 56796e683305SStefano Zampini PetscInt lrows; 56806e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 56816e683305SStefano Zampini if (coarse_mat_is) { 56826e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 56836e683305SStefano Zampini } else { 56846e683305SStefano Zampini lrows = 0; 56856e683305SStefano Zampini } 56866e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 56876e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 56886e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 56896e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56906e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56916e683305SStefano Zampini } 56926e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 56936e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5694c8587f34SStefano Zampini 5695f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5696f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5697f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5698f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5699f9eb5b7dSStefano Zampini } else { 5700f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5701f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5702c8587f34SStefano Zampini } 5703c8587f34SStefano Zampini 57046e683305SStefano Zampini /* print some info if requested */ 57056e683305SStefano Zampini if (pcbddc->dbg_flag) { 57066e683305SStefano Zampini if (!multilevel_allowed) { 57076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57086e683305SStefano Zampini if (multilevel_requested) { 57096e683305SStefano 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); 57106e683305SStefano Zampini } else if (pcbddc->max_levels) { 57116e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 57126e683305SStefano Zampini } 57136e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 57146e683305SStefano Zampini } 57156e683305SStefano Zampini } 57166e683305SStefano Zampini 5717f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 57186e683305SStefano Zampini if (coarse_mat_is) { 57196e683305SStefano Zampini MatReuse coarse_mat_reuse; 57206a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 57216e683305SStefano Zampini if (pcbddc->dbg_flag) { 57226e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 57236e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57246e683305SStefano Zampini } 5725f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5726312be037SStefano Zampini char prefix[256],str_level[16]; 5727e604994aSStefano Zampini size_t len; 57286e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5729422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5730c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5731f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 57325f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5733c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 57346e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5735c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5736c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5737e604994aSStefano Zampini /* prefix */ 5738e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5739e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5740e604994aSStefano Zampini if (!pcbddc->current_level) { 5741e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5742e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5743c8587f34SStefano Zampini } else { 5744e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5745312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5746312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 574734d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5748312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5749e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5750e604994aSStefano Zampini } 5751e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 57523e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 57533e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 57543e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 57553e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5756f9eb5b7dSStefano Zampini /* allow user customization */ 5757f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 57583e3c6dadSStefano Zampini } 57593e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 576051bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 57613e3c6dadSStefano Zampini if (nisdofs) { 57623e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 57633e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 57643e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 57653e3c6dadSStefano Zampini } 57663e3c6dadSStefano Zampini } 57673e3c6dadSStefano Zampini if (nisneu) { 57683e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 57693e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5770312be037SStefano Zampini } 577130368db7SStefano Zampini if (nisvert) { 577230368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 577330368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 577430368db7SStefano Zampini } 5775f9eb5b7dSStefano Zampini 5776f9eb5b7dSStefano Zampini /* get some info after set from options */ 5777f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5778f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 57794f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 57806e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5781f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5782f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5783f9eb5b7dSStefano Zampini } 578439f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 57854f3a063dSStefano Zampini if (isredundant) { 57864f3a063dSStefano Zampini KSP inner_ksp; 57874f3a063dSStefano Zampini PC inner_pc; 57884f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 57894f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 57904f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 57914f3a063dSStefano Zampini } 5792f9eb5b7dSStefano Zampini 5793f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5794fa7f1dd8SStefano Zampini if (coarse_reuse) { 579581d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5796fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 57976e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5798fa7f1dd8SStefano Zampini } else { 57996e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5800fa7f1dd8SStefano Zampini } 5801c8587f34SStefano Zampini if (isbddc || isnn) { 5802720d30f9SStefano Zampini if (isbddc) { /* currently there are no APIs for these options */ 5803720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5804720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 5805720d30f9SStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_saddle_point; 5806720d30f9SStefano Zampini pcbddc_coarse->benign_compute_nonetflux = pcbddc->benign_compute_nonetflux; 5807720d30f9SStefano Zampini if (pcbddc_coarse->benign_compute_nonetflux) { 5808720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 5809720d30f9SStefano Zampini } 5810d4d8cf7bSStefano Zampini } 581122bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 581270cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5813b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 581422b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 58156e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 58166e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 58176e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 58186e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 581922b6e8a2SStefano Zampini } 582070cf5478SStefano Zampini } 582153a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 582270cf5478SStefano Zampini } else { 582322bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 582422bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 582522bc73bbSStefano Zampini } 582622bc73bbSStefano Zampini } else { 58272e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5828c8587f34SStefano Zampini } 5829c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5830c8587f34SStefano Zampini 58313301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 58325a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 58333301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 58343301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 58353301b35fSStefano Zampini } 58363301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 58373301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 58383301b35fSStefano Zampini } 58393301b35fSStefano Zampini if (pc->pmat->spd_set) { 58403301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 58413301b35fSStefano Zampini } 58426e683305SStefano Zampini /* set operators */ 58435f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 58446e683305SStefano Zampini if (pcbddc->dbg_flag) { 58456e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 58466e683305SStefano Zampini } 58476e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 58486e683305SStefano Zampini coarse_mat = 0; 58496e683305SStefano Zampini } 58506e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5851b1ecc7b1SStefano Zampini #if 0 5852b9b85e73SStefano Zampini { 5853b9b85e73SStefano Zampini PetscViewer viewer; 5854b9b85e73SStefano Zampini char filename[256]; 5855b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5856b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 58576a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5858b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5859f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 5860b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5861b9b85e73SStefano Zampini } 5862b9b85e73SStefano Zampini #endif 5863c8587f34SStefano Zampini 5864c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5865298c0119SStefano Zampini #if 0 5866c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5867c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 586898a51de6SStefano Zampini } 5869298c0119SStefano Zampini #endif 5870b0f5fe93SStefano Zampini /* hack */ 587198a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 587298a51de6SStefano Zampini Vec crhs,csol; 587304708bb6SStefano Zampini 5874f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5875f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5876f347579bSStefano Zampini if (!csol) { 58772a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5878f9eb5b7dSStefano Zampini } 5879f347579bSStefano Zampini if (!crhs) { 58802a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5881f347579bSStefano Zampini } 5882b0f5fe93SStefano Zampini } 5883b0f5fe93SStefano Zampini 5884b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5885b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5886b0f5fe93SStefano Zampini 5887b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 58884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58894f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 58904f1b2e48SStefano Zampini } 5891b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5892b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5893b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5894b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5895b0f5fe93SStefano Zampini if (coarse_mat) { 5896b0f5fe93SStefano Zampini Vec nullv; 5897b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5898b0f5fe93SStefano Zampini PetscInt nl; 5899b0f5fe93SStefano Zampini 5900b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5901b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5902b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5903b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5904b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5905b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5906b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5907b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5908b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5909b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5910b0f5fe93SStefano Zampini } 5911b0f5fe93SStefano Zampini } 5912b0f5fe93SStefano Zampini 5913b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5914b0f5fe93SStefano Zampini PetscBool ispreonly; 5915b0f5fe93SStefano Zampini 5916b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5917b0f5fe93SStefano Zampini PetscBool isnull; 5918b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5919d4d8cf7bSStefano Zampini if (0) { 592030368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5921b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5922b0f5fe93SStefano Zampini } else { 5923b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5924b0f5fe93SStefano Zampini } 5925b0f5fe93SStefano Zampini } else { 5926b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5927b0f5fe93SStefano Zampini } 5928b0f5fe93SStefano Zampini } 5929b0f5fe93SStefano Zampini /* setup coarse ksp */ 5930b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5931cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5932cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 59336e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5934c8587f34SStefano Zampini KSP check_ksp; 59352b510759SStefano Zampini KSPType check_ksp_type; 5936c8587f34SStefano Zampini PC check_pc; 59376e683305SStefano Zampini Vec check_vec,coarse_vec; 59386a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 59392b510759SStefano Zampini PetscInt its; 59406e683305SStefano Zampini PetscBool compute_eigs; 59416e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 59426e683305SStefano Zampini PetscInt neigs; 59438e185a42SStefano Zampini const char *prefix; 5944c8587f34SStefano Zampini 59452b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 59466e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5947422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 594823ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5949f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 59502b510759SStefano Zampini if (ispreonly) { 59512b510759SStefano Zampini check_ksp_type = KSPPREONLY; 59526e683305SStefano Zampini compute_eigs = PETSC_FALSE; 59532b510759SStefano Zampini } else { 5954cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 59556e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5956c8587f34SStefano Zampini } 5957c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 59586e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 59596e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 59606e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5961a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5962a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5963a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5964a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5965c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5966c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5967c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5968c8587f34SStefano Zampini /* create random vec */ 59696e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 59706e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5971c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5972c8587f34SStefano Zampini if (CoarseNullSpace) { 5973c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5974c8587f34SStefano Zampini } 59756e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5976c8587f34SStefano Zampini /* solve coarse problem */ 59776e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5978c8587f34SStefano Zampini if (CoarseNullSpace) { 59796e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5980c8587f34SStefano Zampini } 5981cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 59826e683305SStefano Zampini if (compute_eigs) { 5983854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5984854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 59856e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 59866e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 59876e683305SStefano Zampini lambda_min = eigs_r[0]; 59886e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 59896e683305SStefano Zampini if (lambda_max>lambda_min) { 5990cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5991cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5992cbcc2c2aSStefano Zampini } 5993c8587f34SStefano Zampini } 5994c8587f34SStefano Zampini } 5995cbcc2c2aSStefano Zampini 5996c8587f34SStefano Zampini /* check coarse problem residual error */ 59976e683305SStefano Zampini if (pcbddc->dbg_flag) { 59986e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 59996e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 60006e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6001c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 60026e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 60036e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6004c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 6005779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 60066e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 60076e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 60086e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 60096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6010b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6011b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6012b0f5fe93SStefano Zampini } 60136e683305SStefano Zampini if (compute_eigs) { 60146e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6015deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6016c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 60176e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 60186e683305SStefano 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); 60196e683305SStefano Zampini for (i=0;i<neigs;i++) { 60206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6021c8587f34SStefano Zampini } 60226e683305SStefano Zampini } 60236e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 60246e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 60256e683305SStefano Zampini } 6026c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 60276e683305SStefano Zampini if (compute_eigs) { 60286e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 60296e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6030c8587f34SStefano Zampini } 60316e683305SStefano Zampini } 60326e683305SStefano Zampini } 6033cbcc2c2aSStefano Zampini /* print additional info */ 6034cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 60356e683305SStefano Zampini /* waits until all processes reaches this point */ 60366e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6037cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6038cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6039cbcc2c2aSStefano Zampini } 6040cbcc2c2aSStefano Zampini 60412b510759SStefano Zampini /* free memory */ 6042c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6043fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6044c8587f34SStefano Zampini PetscFunctionReturn(0); 6045c8587f34SStefano Zampini } 6046674ae819SStefano Zampini 6047f34684f1SStefano Zampini #undef __FUNCT__ 6048f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6049f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6050f34684f1SStefano Zampini { 6051f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6052f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6053f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6054dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6055dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 605673be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6057dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6058f34684f1SStefano Zampini PetscErrorCode ierr; 6059f34684f1SStefano Zampini 6060f34684f1SStefano Zampini PetscFunctionBegin; 6061f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 60626c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6063dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 60643bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6065dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6066dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6067dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6068dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6069dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6070dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60716c4ed002SBarry 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); 6072dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6073dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6074dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6075dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6076dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6077f34684f1SStefano Zampini 6078f34684f1SStefano Zampini /* check numbering */ 6079f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6080019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6081dc456d91SStefano Zampini PetscInt i; 6082b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6083f34684f1SStefano Zampini 6084f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6085f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6086f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 60871575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6088019a44ceSStefano Zampini /* counter */ 6089019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6090019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6091019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6092019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6093019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6094019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6095f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6096f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6097727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6098f34684f1SStefano Zampini } 6099f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6100f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6101f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6102e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6103e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6104e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6105e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6106f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6107019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6108f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6109019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 611075c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 611175c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6112b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6113019a44ceSStefano 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); 6114f34684f1SStefano Zampini } 6115f34684f1SStefano Zampini } 6116019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6117b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6118f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6119f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6120f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6121f34684f1SStefano Zampini } 6122f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6123f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6124e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6125e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6126f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6127f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6128b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6129ca8b9ea9SStefano Zampini PetscInt *gidxs; 6130ca8b9ea9SStefano Zampini 6131ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 61323bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6133f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6134f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6135f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6136f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 61374bc2dc4bSStefano 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); 6138f34684f1SStefano Zampini } 6139f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6140ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6141f34684f1SStefano Zampini } 6142f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 61431575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6144302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6145f34684f1SStefano Zampini } 61468bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6147f34684f1SStefano Zampini /* get back data */ 6148f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6149f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6150674ae819SStefano Zampini PetscFunctionReturn(0); 6151674ae819SStefano Zampini } 6152674ae819SStefano Zampini 6153e456f2a8SStefano Zampini #undef __FUNCT__ 6154e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6155a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6156e456f2a8SStefano Zampini { 6157e456f2a8SStefano Zampini IS localis_t; 6158a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6159e456f2a8SStefano Zampini PetscScalar *vals; 6160e456f2a8SStefano Zampini PetscErrorCode ierr; 6161e456f2a8SStefano Zampini 6162e456f2a8SStefano Zampini PetscFunctionBegin; 6163a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6164e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6165854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6166e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6167e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6168a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6169a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61701035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6171a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61721035eff8SStefano Zampini } 6173a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6174e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6175e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6176a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6177a7dc3881SStefano Zampini /* now compute set in local ordering */ 6178a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6179a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6180a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6181a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6182a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6183ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6184e456f2a8SStefano Zampini lsize++; 6185e456f2a8SStefano Zampini } 6186e456f2a8SStefano Zampini } 6187854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6188a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6189ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6190e456f2a8SStefano Zampini idxs[lsize++] = i; 6191e456f2a8SStefano Zampini } 6192e456f2a8SStefano Zampini } 6193a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6194a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6195e456f2a8SStefano Zampini *localis = localis_t; 6196e456f2a8SStefano Zampini PetscFunctionReturn(0); 6197e456f2a8SStefano Zampini } 6198906d46d4SStefano Zampini 6199b96c3477SStefano Zampini #undef __FUNCT__ 6200b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 620108122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6202b96c3477SStefano Zampini { 6203a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6204b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6205b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6206a64f4aa4SStefano Zampini Mat S_j; 6207b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6208b96c3477SStefano Zampini PetscBool free_used_adj; 6209b96c3477SStefano Zampini PetscErrorCode ierr; 6210b96c3477SStefano Zampini 6211b96c3477SStefano Zampini PetscFunctionBegin; 6212b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6213b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 621408122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6215b96c3477SStefano Zampini used_xadj = NULL; 6216b96c3477SStefano Zampini used_adjncy = NULL; 6217b96c3477SStefano Zampini } else { 621808122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 621908122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 622008122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 622108122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6222b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6223b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6224b96c3477SStefano Zampini } else { 62252fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6226b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6227b96c3477SStefano Zampini PetscInt nvtxs; 6228b96c3477SStefano Zampini 62292fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 62302fffb893SStefano Zampini if (flg_row) { 6231b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6232b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6233b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6234b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62352fffb893SStefano Zampini } else { 62362fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62372fffb893SStefano Zampini used_xadj = NULL; 62382fffb893SStefano Zampini used_adjncy = NULL; 62392fffb893SStefano Zampini } 62402fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6241b96c3477SStefano Zampini } 6242b96c3477SStefano Zampini } 6243d5574798SStefano Zampini 6244d5574798SStefano Zampini /* setup sub_schurs data */ 6245a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6246df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6247df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6248a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 62499d54b7f4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,NULL,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 6250a64f4aa4SStefano Zampini } else { 62516816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 6252*b7ab4a40SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE;; 6253a3df083aSStefano Zampini PetscInt benign_n; 625472b8c272SStefano Zampini Mat change = NULL; 62559d54b7f4SStefano Zampini Vec scaling = NULL; 625672b8c272SStefano Zampini IS change_primal = NULL; 6257a3df083aSStefano Zampini 62585feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62595feab87aSStefano Zampini PetscInt n_vertices; 62605feab87aSStefano Zampini 62615feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62622034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62635feab87aSStefano Zampini } 626404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 626504708bb6SStefano Zampini if (!isseqaij) { 626604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 626704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 626804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 626904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 627004708bb6SStefano Zampini } else { 6271511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 627204708bb6SStefano Zampini } 627304708bb6SStefano Zampini } 6274a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6275a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6276ca92afb2SStefano Zampini } else { 6277a3df083aSStefano Zampini benign_n = 0; 6278ca92afb2SStefano Zampini } 6279*b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6280*b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6281*b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 628272b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 6283*b7ab4a40SStefano Zampini PetscBool have_loc_change = !!(sub_schurs->change); 6284*b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6285*b7ab4a40SStefano Zampini need_change = !need_change; 6286*b7ab4a40SStefano Zampini } 6287*b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6288*b7ab4a40SStefano 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 */ 6289*b7ab4a40SStefano Zampini if (need_change) { 629088c03ad3SStefano Zampini PC_IS *pcisf; 629188c03ad3SStefano Zampini PC_BDDC *pcbddcf; 629288c03ad3SStefano Zampini PC pcf; 629388c03ad3SStefano Zampini 629488c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 629588c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 629688c03ad3SStefano Zampini } 629788c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 629888c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 629988c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 630088c03ad3SStefano Zampini /* hacks */ 630188c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 630272b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 630372b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 630472b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 630572b8c272SStefano Zampini pcisf->n = pcis->n; 630672b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 630788c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 630888c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 630988c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 631088c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 631188c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 631288c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 631372b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 631488c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 631588c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 631672b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 631772b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 631872b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 631972b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 632072b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 632188c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 632272b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 632388c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 632488c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 632588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 632688c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 632788c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 632888c03ad3SStefano Zampini pcf->ops->destroy = NULL; 632988c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 633088c03ad3SStefano Zampini } 63319d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 63329d54b7f4SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 633372b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 633472b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6335ca92afb2SStefano Zampini } 6336d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6337b96c3477SStefano Zampini 6338b96c3477SStefano Zampini /* free adjacency */ 6339b96c3477SStefano Zampini if (free_used_adj) { 6340b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6341b96c3477SStefano Zampini } 6342b96c3477SStefano Zampini PetscFunctionReturn(0); 6343b96c3477SStefano Zampini } 6344b96c3477SStefano Zampini 6345b96c3477SStefano Zampini #undef __FUNCT__ 6346b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 634708122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6348b96c3477SStefano Zampini { 6349b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6350b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6351b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6352b96c3477SStefano Zampini PCBDDCGraph graph; 6353b96c3477SStefano Zampini PetscErrorCode ierr; 6354b96c3477SStefano Zampini 6355b96c3477SStefano Zampini PetscFunctionBegin; 6356b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 635708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63583301b35fSStefano Zampini IS verticesIS,verticescomm; 63593301b35fSStefano Zampini PetscInt vsize,*idxs; 6360b96c3477SStefano Zampini 6361b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63623301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63633301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63643301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63653301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63663301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6367b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63687fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6369441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63703301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6371b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6372b96c3477SStefano Zampini /* 6373b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6374b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6375b96c3477SStefano Zampini } 6376b96c3477SStefano Zampini */ 6377b96c3477SStefano Zampini } else { 6378b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6379b96c3477SStefano Zampini } 6380b96c3477SStefano Zampini 6381b96c3477SStefano Zampini /* sub_schurs init */ 6382a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6383a64f4aa4SStefano Zampini 6384b96c3477SStefano Zampini /* free graph struct */ 638508122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6386b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6387b96c3477SStefano Zampini } 6388b96c3477SStefano Zampini PetscFunctionReturn(0); 6389b96c3477SStefano Zampini } 6390fa34dd3eSStefano Zampini 6391fa34dd3eSStefano Zampini #undef __FUNCT__ 6392fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6393fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6394fa34dd3eSStefano Zampini { 6395fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6396fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6397fa34dd3eSStefano Zampini PetscErrorCode ierr; 6398fa34dd3eSStefano Zampini 6399fa34dd3eSStefano Zampini PetscFunctionBegin; 6400fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6401fa34dd3eSStefano Zampini IS zerodiag = NULL; 64024f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6403fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 64044f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 640575c01103SStefano Zampini PetscReal norm; 6406fa34dd3eSStefano Zampini PetscInt i; 6407fa34dd3eSStefano Zampini 6408fa34dd3eSStefano Zampini /* B0 and B0_B */ 6409fa34dd3eSStefano Zampini if (zerodiag) { 6410fa34dd3eSStefano Zampini IS dummy; 6411fa34dd3eSStefano Zampini 64124f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 64134f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6414fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6415fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6416fa34dd3eSStefano Zampini } 6417fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6418fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6419fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6420fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6421fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6422fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6423fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6424fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6425fa34dd3eSStefano Zampini /* S_j */ 6426fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6427fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6428fa34dd3eSStefano Zampini 6429fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6430fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6431fa34dd3eSStefano Zampini /* continuous in primal space */ 6432fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);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); 64364f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 64374f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 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->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6443fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6444fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6445fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6446fa34dd3eSStefano Zampini 6447fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6448fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6449fa34dd3eSStefano Zampini /* local with Schur */ 6450fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6451fa34dd3eSStefano Zampini if (zerodiag) { 6452fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64534f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6454fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6455fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6456fa34dd3eSStefano Zampini } 6457fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6458fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6459fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6460fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6461fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6462fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6463fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6464fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6465fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6466fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6467fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6468fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6469fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6470fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6471fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6472fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6473fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6474fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6475fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6476fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6477fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6478fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6479fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6480fa34dd3eSStefano Zampini if (zerodiag) { 6481fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6482fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64834f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6484fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6485fa34dd3eSStefano Zampini } 6486fa34dd3eSStefano Zampini /* BDDC */ 6487fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6488fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6489fa34dd3eSStefano Zampini 6490fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6491fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6492fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6493fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64944f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64954f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6496fa34dd3eSStefano Zampini } 64974f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6498fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6499fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6500fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6501fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6502fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6503fa34dd3eSStefano Zampini } 6504fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6505fa34dd3eSStefano Zampini } 6506