11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5674ae819SStefano Zampini 6a3df083aSStefano Zampini #undef __FUNCT__ 7a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 8a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 9a3df083aSStefano Zampini { 10a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 11a3df083aSStefano Zampini PetscErrorCode ierr; 12a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 13a3df083aSStefano Zampini 14a3df083aSStefano Zampini PetscFunctionBegin; 15a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 16a3df083aSStefano Zampini if (transpose) { 17a3df083aSStefano Zampini apply_right = ctx->apply_left; 18a3df083aSStefano Zampini apply_left = ctx->apply_right; 19a3df083aSStefano Zampini } else { 20a3df083aSStefano Zampini apply_right = ctx->apply_right; 21a3df083aSStefano Zampini apply_left = ctx->apply_left; 22a3df083aSStefano Zampini } 23a3df083aSStefano Zampini reset_x = PETSC_FALSE; 24a3df083aSStefano Zampini if (apply_right) { 25a3df083aSStefano Zampini const PetscScalar *ax; 26a3df083aSStefano Zampini PetscInt nl,i; 27a3df083aSStefano Zampini 28a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 29a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 30a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 31a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 32a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 33a3df083aSStefano Zampini PetscScalar sum,val; 34a3df083aSStefano Zampini const PetscInt *idxs; 35a3df083aSStefano Zampini PetscInt nz,j; 36a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 37a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 38a3df083aSStefano Zampini sum = 0.; 39a3df083aSStefano Zampini if (ctx->apply_p0) { 40a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 41a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 42a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 43a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 44a3df083aSStefano Zampini } 45a3df083aSStefano Zampini } else { 46a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 47a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 48a3df083aSStefano Zampini } 49a3df083aSStefano Zampini } 50a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 51a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 52a3df083aSStefano Zampini } 53a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 54a3df083aSStefano Zampini reset_x = PETSC_TRUE; 55a3df083aSStefano Zampini } 56a3df083aSStefano Zampini if (transpose) { 57a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 58a3df083aSStefano Zampini } else { 59a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 60a3df083aSStefano Zampini } 61a3df083aSStefano Zampini if (reset_x) { 62a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 63a3df083aSStefano Zampini } 64a3df083aSStefano Zampini if (apply_left) { 65a3df083aSStefano Zampini PetscScalar *ay; 66a3df083aSStefano Zampini PetscInt i; 67a3df083aSStefano Zampini 68a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 69a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 70a3df083aSStefano Zampini PetscScalar sum,val; 71a3df083aSStefano Zampini const PetscInt *idxs; 72a3df083aSStefano Zampini PetscInt nz,j; 73a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 74a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 75a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 76a3df083aSStefano Zampini if (ctx->apply_p0) { 77a3df083aSStefano Zampini sum = 0.; 78a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 79a3df083aSStefano Zampini sum += ay[idxs[j]]; 80a3df083aSStefano Zampini ay[idxs[j]] += val; 81a3df083aSStefano Zampini } 82a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 83a3df083aSStefano Zampini } else { 84a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 85a3df083aSStefano Zampini ay[idxs[j]] += val; 86a3df083aSStefano Zampini } 87a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 88a3df083aSStefano Zampini } 89a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 90a3df083aSStefano Zampini } 91a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 92a3df083aSStefano Zampini } 93a3df083aSStefano Zampini PetscFunctionReturn(0); 94a3df083aSStefano Zampini } 95a3df083aSStefano Zampini 96a3df083aSStefano Zampini #undef __FUNCT__ 97a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 98a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 99a3df083aSStefano Zampini { 100a3df083aSStefano Zampini PetscErrorCode ierr; 101a3df083aSStefano Zampini 102a3df083aSStefano Zampini PetscFunctionBegin; 103a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 104a3df083aSStefano Zampini PetscFunctionReturn(0); 105a3df083aSStefano Zampini } 106a3df083aSStefano Zampini 107a3df083aSStefano Zampini #undef __FUNCT__ 108a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 109a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 110a3df083aSStefano Zampini { 111a3df083aSStefano Zampini PetscErrorCode ierr; 112a3df083aSStefano Zampini 113a3df083aSStefano Zampini PetscFunctionBegin; 114a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 115a3df083aSStefano Zampini PetscFunctionReturn(0); 116a3df083aSStefano Zampini } 117a3df083aSStefano Zampini 118a3df083aSStefano Zampini #undef __FUNCT__ 119a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 120a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 121a3df083aSStefano Zampini { 122a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 123a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 124a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 125a3df083aSStefano Zampini PetscErrorCode ierr; 126a3df083aSStefano Zampini 127a3df083aSStefano Zampini PetscFunctionBegin; 128a3df083aSStefano Zampini if (!restore) { 1291dd7afcfSStefano Zampini Mat A_IB,A_BI; 130a3df083aSStefano Zampini PetscScalar *work; 131df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 132a3df083aSStefano Zampini 1331dd7afcfSStefano Zampini if (pcbddc->benign_original_mat) { 1341dd7afcfSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored"); 1351dd7afcfSStefano Zampini } 1361dd7afcfSStefano Zampini if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) { 1371dd7afcfSStefano Zampini PetscFunctionReturn(0); 1381dd7afcfSStefano Zampini } 139a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 140a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 141a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 142a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 145a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 146a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 147a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 148a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 149a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 150a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 151059032f7SStefano Zampini if (reuse) { 152a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 1531dd7afcfSStefano Zampini ctx->free = PETSC_FALSE; 154059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 155059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 156059032f7SStefano Zampini PetscInt i; 157059032f7SStefano Zampini 158059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 159059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 160059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 161059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 162059032f7SStefano Zampini } 163059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 1641dd7afcfSStefano Zampini ctx->free = PETSC_TRUE; 165059032f7SStefano Zampini } 166a3df083aSStefano Zampini ctx->A = pcis->A_IB; 167a3df083aSStefano Zampini ctx->work = work; 168a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 169a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 170a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 171a3df083aSStefano Zampini pcis->A_IB = A_IB; 172a3df083aSStefano Zampini 173a3df083aSStefano Zampini /* A_BI as A_IB^T */ 174a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 175a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 176a3df083aSStefano Zampini pcis->A_BI = A_BI; 177a3df083aSStefano Zampini } else { 1781dd7afcfSStefano Zampini if (!pcbddc->benign_original_mat) { 1791dd7afcfSStefano Zampini PetscFunctionReturn(0); 1801dd7afcfSStefano Zampini } 181a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 182a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 183a3df083aSStefano Zampini pcis->A_IB = ctx->A; 1841dd7afcfSStefano Zampini ctx->A = NULL; 1851dd7afcfSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 1861dd7afcfSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 1871dd7afcfSStefano Zampini pcbddc->benign_original_mat = NULL; 1881dd7afcfSStefano Zampini if (ctx->free) { 189059032f7SStefano Zampini PetscInt i; 1901dd7afcfSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 191059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 192059032f7SStefano Zampini } 193059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 194059032f7SStefano Zampini } 195a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 196a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 197a3df083aSStefano Zampini } 198a3df083aSStefano Zampini PetscFunctionReturn(0); 199a3df083aSStefano Zampini } 200a3df083aSStefano Zampini 201a3df083aSStefano Zampini /* used just in bddc debug mode */ 202a3df083aSStefano Zampini #undef __FUNCT__ 203a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 204a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 205a3df083aSStefano Zampini { 206a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 207a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 208a3df083aSStefano Zampini Mat An; 209a3df083aSStefano Zampini PetscErrorCode ierr; 210a3df083aSStefano Zampini 211a3df083aSStefano Zampini PetscFunctionBegin; 212a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 213a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 214a3df083aSStefano Zampini if (is1) { 215a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 216a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 217a3df083aSStefano Zampini } else { 218a3df083aSStefano Zampini *B = An; 219a3df083aSStefano Zampini } 220a3df083aSStefano Zampini PetscFunctionReturn(0); 221a3df083aSStefano Zampini } 222a3df083aSStefano Zampini 2231cf9b237SStefano Zampini /* TODO: add reuse flag */ 2241cf9b237SStefano Zampini #undef __FUNCT__ 2251cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2261cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2271cf9b237SStefano Zampini { 2281cf9b237SStefano Zampini Mat Bt; 2291cf9b237SStefano Zampini PetscScalar *a,*bdata; 2301cf9b237SStefano Zampini const PetscInt *ii,*ij; 2311cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2321cf9b237SStefano Zampini PetscBool flg_row; 2331cf9b237SStefano Zampini PetscErrorCode ierr; 2341cf9b237SStefano Zampini 2351cf9b237SStefano Zampini PetscFunctionBegin; 2361cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2371cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2381cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2391cf9b237SStefano Zampini nnz = n; 2401cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2411cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2421cf9b237SStefano Zampini } 2431cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2441cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2451cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2461cf9b237SStefano Zampini nnz = 0; 2471cf9b237SStefano Zampini bii[0] = 0; 2481cf9b237SStefano Zampini for (i=0;i<n;i++) { 2491cf9b237SStefano Zampini PetscInt j; 2501cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2511cf9b237SStefano Zampini PetscScalar entry = a[j]; 2521cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2531cf9b237SStefano Zampini bij[nnz] = ij[j]; 2541cf9b237SStefano Zampini bdata[nnz] = entry; 2551cf9b237SStefano Zampini nnz++; 2561cf9b237SStefano Zampini } 2571cf9b237SStefano Zampini } 2581cf9b237SStefano Zampini bii[i+1] = nnz; 2591cf9b237SStefano Zampini } 2601cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2611cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2621cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2631cf9b237SStefano Zampini { 2641cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2651cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2661cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2671cf9b237SStefano Zampini } 2681cf9b237SStefano Zampini *B = Bt; 2691cf9b237SStefano Zampini PetscFunctionReturn(0); 2701cf9b237SStefano Zampini } 2711cf9b237SStefano Zampini 272674ae819SStefano Zampini #undef __FUNCT__ 2734f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2744f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2754f1b2e48SStefano Zampini { 2764f1b2e48SStefano Zampini Mat B; 2774f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2784f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2794f1b2e48SStefano Zampini PCBDDCGraph graph; 2804f1b2e48SStefano Zampini PetscInt i,n; 2814f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 2824f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 2834f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 2844f1b2e48SStefano Zampini PetscErrorCode ierr; 2854f1b2e48SStefano Zampini 2864f1b2e48SStefano Zampini PetscFunctionBegin; 28763c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 28863c961adSStefano Zampini *ncc = 0; 28963c961adSStefano Zampini *cc = NULL; 29063c961adSStefano Zampini PetscFunctionReturn(0); 29163c961adSStefano Zampini } 2924f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 2934f1b2e48SStefano Zampini if (!isseqaij && filter) { 2941cf9b237SStefano Zampini PetscBool isseqdense; 2951cf9b237SStefano Zampini 2961cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 2971cf9b237SStefano Zampini if (!isseqdense) { 2984f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 2991cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3001cf9b237SStefano Zampini PetscScalar *array; 3011cf9b237SStefano Zampini PetscReal chop=1.e-6; 3021cf9b237SStefano Zampini 3031cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3041cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3051cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3061cf9b237SStefano Zampini for (i=0;i<n;i++) { 3071cf9b237SStefano Zampini PetscInt j; 3081cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3091cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3101cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3111cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3121cf9b237SStefano Zampini } 3131cf9b237SStefano Zampini } 3141cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3159d54b7f4SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr); 3161cf9b237SStefano Zampini } 3174f1b2e48SStefano Zampini } else { 3184f1b2e48SStefano Zampini B = A; 3194f1b2e48SStefano Zampini } 3204f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3214f1b2e48SStefano Zampini 3224f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3234f1b2e48SStefano Zampini if (filter) { 3244f1b2e48SStefano Zampini PetscScalar *data; 3254f1b2e48SStefano Zampini PetscInt j,cum; 3264f1b2e48SStefano Zampini 3274f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3284f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3294f1b2e48SStefano Zampini cum = 0; 3304f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3314f1b2e48SStefano Zampini PetscInt t; 3324f1b2e48SStefano Zampini 3334f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3344f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3354f1b2e48SStefano Zampini continue; 3364f1b2e48SStefano Zampini } 3374f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini t = xadj_filtered[i]; 3404f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3414f1b2e48SStefano Zampini cum += t; 3424f1b2e48SStefano Zampini } 3434f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3444f1b2e48SStefano Zampini } else { 3454f1b2e48SStefano Zampini xadj_filtered = NULL; 3464f1b2e48SStefano Zampini adjncy_filtered = NULL; 3474f1b2e48SStefano Zampini } 3484f1b2e48SStefano Zampini 3494f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3504f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3514f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3524f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3534f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3544f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3554f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3564f1b2e48SStefano Zampini if (xadj_filtered) { 3574f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3584f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3594f1b2e48SStefano Zampini } else { 3604f1b2e48SStefano Zampini graph->xadj = xadj; 3614f1b2e48SStefano Zampini graph->adjncy = adjncy; 3624f1b2e48SStefano Zampini } 3634f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3644f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3654f1b2e48SStefano Zampini /* partial clean up */ 3664f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3674f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3681cf9b237SStefano Zampini if (A != B) { 3694f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3704f1b2e48SStefano Zampini } 3714f1b2e48SStefano Zampini 3724f1b2e48SStefano Zampini /* get back data */ 3731cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3741cf9b237SStefano Zampini if (cc) { 3754f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3764f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3774f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);CHKERRQ(ierr); 3784f1b2e48SStefano Zampini } 3794f1b2e48SStefano Zampini *cc = cc_n; 3801cf9b237SStefano Zampini } 3814f1b2e48SStefano Zampini /* clean up graph */ 3824f1b2e48SStefano Zampini graph->xadj = 0; 3834f1b2e48SStefano Zampini graph->adjncy = 0; 3844f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 3854f1b2e48SStefano Zampini PetscFunctionReturn(0); 3864f1b2e48SStefano Zampini } 3874f1b2e48SStefano Zampini 3884f1b2e48SStefano Zampini #undef __FUNCT__ 3895408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 3905408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 3915408967cSStefano Zampini { 3925408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3935408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 394dee84bffSStefano Zampini IS dirIS = NULL; 3954f1b2e48SStefano Zampini PetscInt i; 3965408967cSStefano Zampini PetscErrorCode ierr; 3975408967cSStefano Zampini 3985408967cSStefano Zampini PetscFunctionBegin; 399dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4005408967cSStefano Zampini if (zerodiag) { 4015408967cSStefano Zampini Mat A; 4025408967cSStefano Zampini Vec vec3_N; 4035408967cSStefano Zampini PetscScalar *vals; 4045408967cSStefano Zampini const PetscInt *idxs; 405d12d3064SStefano Zampini PetscInt nz,*count; 4065408967cSStefano Zampini 4075408967cSStefano Zampini /* p0 */ 4085408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4095408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4105408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4115408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4124f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4135408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4145408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4155408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4165408967cSStefano Zampini /* v_I */ 4175408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4185408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4195408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4205408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4215408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4225408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4235408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4245408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4255408967cSStefano Zampini if (dirIS) { 4265408967cSStefano Zampini PetscInt n; 4275408967cSStefano Zampini 4285408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4295408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4305408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4315408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4325408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4335408967cSStefano Zampini } 4345408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4355408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4365408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4375408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4385408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 441*b9b0e38cSStefano Zampini if (!pcbddc->current_level && PetscAbsScalar(vals[0]) > PETSC_SMALL) { 442*b9b0e38cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %1.6e (should be numerically 0.)",PetscAbsScalar(vals[0])); 4435408967cSStefano Zampini } 4445408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4455408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 446d12d3064SStefano Zampini 447d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 448d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 449d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 450d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 451d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 452d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 453d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 454d12d3064SStefano Zampini if (count[idxs[i]]) { 455d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 456d12d3064SStefano Zampini } 457d4d8cf7bSStefano Zampini } 458d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 459d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4605408967cSStefano Zampini } 461dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4625408967cSStefano Zampini 4635408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4645408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4665408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4674f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4685408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4694f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4704f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4714f1b2e48SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %1.4e at %d instead of %1.4e\n",pcbddc->benign_p0[i],i,-PetscGlobalRank-i);CHKERRQ(ierr); 4724f1b2e48SStefano Zampini } 4735408967cSStefano Zampini } 4745408967cSStefano Zampini PetscFunctionReturn(0); 4755408967cSStefano Zampini } 4765408967cSStefano Zampini 4775408967cSStefano Zampini #undef __FUNCT__ 478339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 479339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 480339f8db1SStefano Zampini { 481339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4824f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 483b0f5fe93SStefano Zampini PetscInt nz,n; 4844f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 485339f8db1SStefano Zampini PetscErrorCode ierr; 486339f8db1SStefano Zampini 487339f8db1SStefano Zampini PetscFunctionBegin; 4889f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4899f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 490a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 491a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 492a3df083aSStefano Zampini } 493a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 494a3df083aSStefano Zampini pcbddc->benign_n = 0; 4954f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4964f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 4974f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 4984f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 4994f1b2e48SStefano Zampini since the local Schur complements are SPD 5004f1b2e48SStefano Zampini */ 5014f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5024f1b2e48SStefano Zampini have_null = PETSC_TRUE; 50340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5044f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5054f1b2e48SStefano Zampini 5064f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5074f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5084f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5094f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 510ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 51140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 51240fa8d13SStefano Zampini if (!sorted) { 51340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 51440fa8d13SStefano Zampini } 51540fa8d13SStefano Zampini } else { 51640fa8d13SStefano Zampini pressures = NULL; 51740fa8d13SStefano Zampini } 51897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 51997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 52097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 521339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 522339f8db1SStefano Zampini if (!sorted) { 523339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 524339f8db1SStefano Zampini } 525339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5264f1b2e48SStefano Zampini if (!nz) { 5274f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5284f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 52940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 53040fa8d13SStefano Zampini } 5314f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5324f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5334f1b2e48SStefano Zampini zerodiag_subs = NULL; 5344f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5354f1b2e48SStefano Zampini if (has_null_pressures) { 5364f1b2e48SStefano Zampini IS *subs; 5374f1b2e48SStefano Zampini PetscInt nsubs,i; 5384f1b2e48SStefano Zampini 5394f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5404f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5414f1b2e48SStefano Zampini if (nsubs > 1) { 5424f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5444f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5454f1b2e48SStefano Zampini IS t_zerodiag_subs; 5464f1b2e48SStefano Zampini PetscInt nl; 5474f1b2e48SStefano Zampini 5484f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5494f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5504f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5514f1b2e48SStefano Zampini if (nl) { 5524f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5534f1b2e48SStefano Zampini 5544f1b2e48SStefano Zampini if (pressures) { 5554f1b2e48SStefano Zampini IS t_pressure_subs; 5564f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5574f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5584f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5594f1b2e48SStefano Zampini } 5604f1b2e48SStefano Zampini if (valid) { 5614f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini pcbddc->benign_n++; 5634f1b2e48SStefano Zampini } else { 5644f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5654f1b2e48SStefano Zampini } 5664f1b2e48SStefano Zampini } 5674f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini } 5704f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5714f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5724f1b2e48SStefano Zampini if (pressures) { 5734f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5744f1b2e48SStefano Zampini } 5754f1b2e48SStefano Zampini if (valid) { 5764f1b2e48SStefano Zampini pcbddc->benign_n = 1; 577ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5784f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5794f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5804f1b2e48SStefano Zampini } 5814f1b2e48SStefano Zampini } 5824f1b2e48SStefano Zampini } 5834f1b2e48SStefano Zampini 5844f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 585*b9b0e38cSStefano Zampini PetscInt n; 586*b9b0e38cSStefano Zampini 5874f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 589*b9b0e38cSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 590*b9b0e38cSStefano Zampini if (n) { 5914f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5924f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5934f1b2e48SStefano Zampini } 594*b9b0e38cSStefano Zampini } 5954f1b2e48SStefano Zampini 5964f1b2e48SStefano Zampini /* final check for null pressures */ 5974f1b2e48SStefano Zampini if (zerodiag && pressures) { 5984f1b2e48SStefano Zampini PetscInt nz,np; 5994f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6004f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 6014f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 6024f1b2e48SStefano Zampini } 6034f1b2e48SStefano Zampini 6044f1b2e48SStefano Zampini if (recompute_zerodiag) { 6054f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6074f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6084f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6094f1b2e48SStefano Zampini } else { 6104f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6114f1b2e48SStefano Zampini 6124f1b2e48SStefano Zampini nzn = 0; 6134f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6144f1b2e48SStefano Zampini PetscInt ns; 6154f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6164f1b2e48SStefano Zampini nzn += ns; 6174f1b2e48SStefano Zampini } 6184f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6194f1b2e48SStefano Zampini nzn = 0; 6204f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6214f1b2e48SStefano Zampini PetscInt ns,*idxs; 6224f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6234f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6244f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6254f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6264f1b2e48SStefano Zampini nzn += ns; 6274f1b2e48SStefano Zampini } 6284f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6294f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6304f1b2e48SStefano Zampini } 6314f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6324f1b2e48SStefano Zampini } 6334f1b2e48SStefano Zampini 634b3afcdbeSStefano Zampini /* no-net-flux */ 635b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 636b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 637b3afcdbeSStefano Zampini MatNullSpace near_null_space; 638b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 639b3afcdbeSStefano Zampini Vec quad_vec; 640b3afcdbeSStefano Zampini PetscScalar *pvals; 641b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 642b3afcdbeSStefano Zampini 643b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 644b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 645b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 646b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 647b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 648b3afcdbeSStefano Zampini 649b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 650b3afcdbeSStefano Zampini np = 0; 651b3afcdbeSStefano Zampini if (zerodiag) { 652b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 653b3afcdbeSStefano Zampini } else if (pressures) { 654b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 655b3afcdbeSStefano Zampini } 656b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 657b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 658b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 659b3afcdbeSStefano Zampini if (np) { 660b3afcdbeSStefano Zampini const PetscInt *pidxs; 661b3afcdbeSStefano Zampini 662b3afcdbeSStefano Zampini if (zerodiag) { 663b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 664b3afcdbeSStefano Zampini } else if (pressures) { 665b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 666b3afcdbeSStefano Zampini } 667b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 668b3afcdbeSStefano Zampini if (zerodiag) { 669b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 670b3afcdbeSStefano Zampini } else if (pressures) { 671b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 672b3afcdbeSStefano Zampini } 673b3afcdbeSStefano Zampini } 674b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 675b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 676b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 677b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 678b3afcdbeSStefano Zampini 679b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 680b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 681b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = -1; 682b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 683b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 684b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 685b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 686b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 687b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 688b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) { 689b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 690b3afcdbeSStefano Zampini pvals[i] = 0.; 691b3afcdbeSStefano Zampini } 692b3afcdbeSStefano Zampini } 693b3afcdbeSStefano Zampini ierr = PetscMalloc1(matis->sf_nleaves,&dummyins);CHKERRQ(ierr); 694b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) dummyins[i] = i; 695b3afcdbeSStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf_nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 696b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 697b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 698b3afcdbeSStefano Zampini 699b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 700b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 701b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 702b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 703b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 704b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 705b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 706b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 707b3afcdbeSStefano Zampini } 708b3afcdbeSStefano Zampini 709b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7104f1b2e48SStefano Zampini if (has_null_pressures) { 7114f1b2e48SStefano Zampini IS zerodiagc; 7124f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7134f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7144f1b2e48SStefano Zampini 7154f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 716339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 717339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 718339f8db1SStefano Zampini /* local change of basis for pressures */ 719339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 72097d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 721339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 722339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 723339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7244f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7264f1b2e48SStefano Zampini PetscInt nzs,j; 7274f1b2e48SStefano Zampini 7284f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7294f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7304f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7314f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7324f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7334f1b2e48SStefano Zampini } 734339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 735339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 736339f8db1SStefano Zampini /* set identity on velocities */ 737339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 738339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 739339f8db1SStefano Zampini } 7404f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7414f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7429f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7434f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 744339f8db1SStefano Zampini /* set change on pressures */ 7454f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7464f1b2e48SStefano Zampini PetscScalar *array; 7474f1b2e48SStefano Zampini PetscInt nzs; 7484f1b2e48SStefano Zampini 7494f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7504f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7514f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 752339f8db1SStefano Zampini PetscScalar vals[2]; 753339f8db1SStefano Zampini PetscInt cols[2]; 754339f8db1SStefano Zampini 755339f8db1SStefano Zampini cols[0] = idxs[i]; 7564f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 757339f8db1SStefano Zampini vals[0] = 1.; 758b0f5fe93SStefano Zampini vals[1] = 1.; 7594f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 760339f8db1SStefano Zampini } 7614f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7624f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7634f1b2e48SStefano Zampini array[nzs-1] = 1.; 7644f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7654f1b2e48SStefano Zampini /* store local idxs for p0 */ 7664f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7674f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 768339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7694f1b2e48SStefano Zampini } 770339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 771339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 772a3df083aSStefano Zampini /* project if needed */ 773a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7741dd7afcfSStefano Zampini Mat M; 7751dd7afcfSStefano Zampini 7761dd7afcfSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 777339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 7781dd7afcfSStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 7791dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 780a3df083aSStefano Zampini } 7814f1b2e48SStefano Zampini /* store global idxs for p0 */ 7824f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 783339f8db1SStefano Zampini } 784ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7854f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 786b0f5fe93SStefano Zampini 787b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 788b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 789339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 790339f8db1SStefano Zampini PetscFunctionReturn(0); 791339f8db1SStefano Zampini } 792339f8db1SStefano Zampini 793339f8db1SStefano Zampini #undef __FUNCT__ 794015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 795015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 796efc2fbd9SStefano Zampini { 797efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 798efc2fbd9SStefano Zampini PetscErrorCode ierr; 799efc2fbd9SStefano Zampini 800efc2fbd9SStefano Zampini PetscFunctionBegin; 801efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 802efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 8034f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 804efc2fbd9SStefano Zampini } 805015636ebSStefano Zampini if (get) { /* use SF to get values */ 806efc2fbd9SStefano Zampini PetscScalar *array; 807efc2fbd9SStefano Zampini 808efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8094f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8104f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 811efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8124f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 8134f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 814efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 815efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 816efc2fbd9SStefano Zampini } 817efc2fbd9SStefano Zampini PetscFunctionReturn(0); 818efc2fbd9SStefano Zampini } 819efc2fbd9SStefano Zampini 820efc2fbd9SStefano Zampini #undef __FUNCT__ 821c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 822c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 823c263805aSStefano Zampini { 824c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 825c263805aSStefano Zampini PetscErrorCode ierr; 826c263805aSStefano Zampini 827c263805aSStefano Zampini PetscFunctionBegin; 828c263805aSStefano Zampini /* TODO: add error checking 829c263805aSStefano Zampini - avoid nested pop (or push) calls. 830c263805aSStefano Zampini - cannot push before pop. 8311c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 832c263805aSStefano Zampini */ 8334f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 834efc2fbd9SStefano Zampini PetscFunctionReturn(0); 835efc2fbd9SStefano Zampini } 836c263805aSStefano Zampini if (pop) { 837a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8384f1b2e48SStefano Zampini IS is_p0; 8394f1b2e48SStefano Zampini MatReuse reuse; 840c263805aSStefano Zampini 841c263805aSStefano Zampini /* extract B_0 */ 8424f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8434f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8444f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8454f1b2e48SStefano Zampini } 8464f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8474f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 848c263805aSStefano Zampini /* remove rows and cols from local problem */ 849c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 85097d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8514f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8524f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 853a3df083aSStefano Zampini } else { 854a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 855a3df083aSStefano Zampini PetscScalar *vals; 856a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 857a3df083aSStefano Zampini 858a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 859a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 860a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8610b5adadeSStefano Zampini PetscInt *nnz; 862a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 863a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 864a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 865331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 866331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 867331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 868331e053bSStefano Zampini nnz[i] = n - nnz[i]; 869331e053bSStefano Zampini } 870331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 871331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 872331e053bSStefano Zampini } 873a3df083aSStefano Zampini 874a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 875a3df083aSStefano Zampini PetscScalar *array; 876a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 877a3df083aSStefano Zampini 878a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 879a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 880a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 881a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 882a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 883a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 884a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 885a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 886a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 887a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 888a3df083aSStefano Zampini cum = 0; 889a3df083aSStefano Zampini for (j=0;j<n;j++) { 890a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 891a3df083aSStefano Zampini vals[cum] = array[j]; 892a3df083aSStefano Zampini idxs_ins[cum] = j; 893a3df083aSStefano Zampini cum++; 894a3df083aSStefano Zampini } 895a3df083aSStefano Zampini } 896a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 897a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 898a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 899a3df083aSStefano Zampini } 900a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 901a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 902a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 903a3df083aSStefano Zampini } 904c263805aSStefano Zampini } else { /* push */ 905a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9064f1b2e48SStefano Zampini PetscInt i; 9074f1b2e48SStefano Zampini 9084f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9094f1b2e48SStefano Zampini PetscScalar *B0_vals; 9104f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9114f1b2e48SStefano Zampini 9124f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9134f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9147b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9154f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9164f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9174f1b2e48SStefano Zampini } 918c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 919c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 920a3df083aSStefano Zampini } else { 921a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 922a3df083aSStefano Zampini } 923c263805aSStefano Zampini } 924c263805aSStefano Zampini PetscFunctionReturn(0); 925c263805aSStefano Zampini } 926c263805aSStefano Zampini 927c263805aSStefano Zampini #undef __FUNCT__ 928b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 92908122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 930b1b3d7a2SStefano Zampini { 931b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 93208122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 93308122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 93408122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 93508122e43SStefano Zampini PetscScalar *work,lwork; 93608122e43SStefano Zampini PetscScalar *St,*S,*eigv; 93708122e43SStefano Zampini PetscScalar *Sarray,*Starray; 93808122e43SStefano Zampini PetscReal *eigs,thresh; 9391b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 940f6f667cfSStefano Zampini PetscBool allocated_S_St; 94108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 94208122e43SStefano Zampini PetscReal *rwork; 94308122e43SStefano Zampini #endif 944b1b3d7a2SStefano Zampini PetscErrorCode ierr; 945b1b3d7a2SStefano Zampini 946b1b3d7a2SStefano Zampini PetscFunctionBegin; 947af25d912SStefano Zampini if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 948af25d912SStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 94906a4e24aSStefano Zampini 950fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 951fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 952fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 953fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9541575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 955fd14bc51SStefano Zampini } 956fd14bc51SStefano Zampini 957e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 958e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 959e496cd5dSStefano Zampini } 960e496cd5dSStefano Zampini 96108122e43SStefano Zampini /* max size of subsets */ 96208122e43SStefano Zampini mss = 0; 96308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 96408122e43SStefano Zampini PetscInt subset_size; 965862806e4SStefano Zampini 96608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 96708122e43SStefano Zampini mss = PetscMax(mss,subset_size); 96808122e43SStefano Zampini } 96908122e43SStefano Zampini 97008122e43SStefano Zampini /* min/max and threshold */ 97108122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 972f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 97308122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 974f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 975f6f667cfSStefano Zampini if (nmin) { 976f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 977f6f667cfSStefano Zampini } 97808122e43SStefano Zampini 97908122e43SStefano Zampini /* allocate lapack workspace */ 98008122e43SStefano Zampini cum = cum2 = 0; 98108122e43SStefano Zampini maxneigs = 0; 98208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 98308122e43SStefano Zampini PetscInt n,subset_size; 984f6f667cfSStefano Zampini 98508122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 98608122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9879162d606SStefano Zampini cum += subset_size; 9889162d606SStefano Zampini cum2 += subset_size*n; 98908122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 99008122e43SStefano Zampini } 99108122e43SStefano Zampini if (mss) { 9929ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 99308122e43SStefano Zampini PetscBLASInt B_itype = 1; 99408122e43SStefano Zampini PetscBLASInt B_N = mss; 9954c6709b3SStefano Zampini PetscReal zero = 0.0; 9964c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 99708122e43SStefano Zampini 99808122e43SStefano Zampini B_lwork = -1; 99908122e43SStefano Zampini S = NULL; 100008122e43SStefano Zampini St = NULL; 1001a58a30b4SStefano Zampini eigs = NULL; 1002a58a30b4SStefano Zampini eigv = NULL; 1003a58a30b4SStefano Zampini B_iwork = NULL; 1004a58a30b4SStefano Zampini B_ifail = NULL; 1005d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1006d1710679SStefano Zampini rwork = NULL; 1007d1710679SStefano Zampini #endif 10088bec7fa6SStefano Zampini thresh = 1.0; 100908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 101008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 101108122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 101208122e43SStefano Zampini #else 101308122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 101408122e43SStefano Zampini #endif 101508122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 101608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 101708122e43SStefano Zampini } else { 101808122e43SStefano Zampini /* TODO */ 101908122e43SStefano Zampini } 102008122e43SStefano Zampini } else { 102108122e43SStefano Zampini lwork = 0; 102208122e43SStefano Zampini } 102308122e43SStefano Zampini 102408122e43SStefano Zampini nv = 0; 1025d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 1026d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 102708122e43SStefano Zampini } 10284c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1029f6f667cfSStefano Zampini if (allocated_S_St) { 1030f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1031f6f667cfSStefano Zampini } 1032f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 103308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 103408122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 103508122e43SStefano Zampini #endif 10369162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10379162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10389162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 103908122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10409162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 104108122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 104208122e43SStefano Zampini 104308122e43SStefano Zampini maxneigs = 0; 104472b8c272SStefano Zampini cum = cumarray = 0; 10459162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10469162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1047d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 104808122e43SStefano Zampini const PetscInt *idxs; 104908122e43SStefano Zampini 1050d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 105108122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 105208122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 105308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 105408122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10559162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10569162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 105708122e43SStefano Zampini } 1058d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 105908122e43SStefano Zampini } 106008122e43SStefano Zampini 106108122e43SStefano Zampini if (mss) { /* multilevel */ 106208122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 106308122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 106408122e43SStefano Zampini } 106508122e43SStefano Zampini 1066ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 106708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 106808122e43SStefano Zampini const PetscInt *idxs; 10699d54b7f4SStefano Zampini PetscReal upper,lower; 1070862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 107108122e43SStefano Zampini PetscBLASInt B_N; 1072aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 107308122e43SStefano Zampini 10749d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 10759d54b7f4SStefano Zampini upper = PETSC_MAX_REAL; 10769d54b7f4SStefano Zampini lower = thresh; 10779d54b7f4SStefano Zampini } else { 10789d54b7f4SStefano Zampini upper = 1./thresh; 10799d54b7f4SStefano Zampini lower = 0.; 10809d54b7f4SStefano Zampini } 1081862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1082ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1083f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1084f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10859ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1086aff50787SStefano Zampini PetscInt j,k; 1087aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1088aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1089aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109008122e43SStefano Zampini } 109108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1092aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1093aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1094aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1095aff50787SStefano Zampini } 109608122e43SStefano Zampini } 109708122e43SStefano Zampini } else { 109808122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 109908122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110008122e43SStefano Zampini } 11018bec7fa6SStefano Zampini } else { 1102f6f667cfSStefano Zampini S = Sarray + cumarray; 1103f6f667cfSStefano Zampini St = Starray + cumarray; 11048bec7fa6SStefano Zampini } 1105aff50787SStefano Zampini /* see if we can save some work */ 1106b7ab4a40SStefano Zampini if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) { 1107aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1108aff50787SStefano Zampini } 1109aff50787SStefano Zampini 1110b7ab4a40SStefano Zampini if (same_data && !sub_schurs->change) { /* there's no need of constraints here */ 1111aff50787SStefano Zampini B_neigs = 0; 1112aff50787SStefano Zampini } else { 11139ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 111408122e43SStefano Zampini PetscBLASInt B_itype = 1; 1115f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11164c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11179552c7c7SStefano Zampini PetscInt nmin_s; 1118b7ab4a40SStefano Zampini PetscBool compute_range = PETSC_FALSE; 111908122e43SStefano Zampini 1120fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11218bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 1122fd14bc51SStefano Zampini } 1123d16cbb6bSStefano Zampini 1124b7ab4a40SStefano Zampini compute_range = PETSC_FALSE; 1125b7ab4a40SStefano Zampini if (thresh > 1.+PETSC_SMALL && !same_data) { 1126b7ab4a40SStefano Zampini compute_range = PETSC_TRUE; 1127b7ab4a40SStefano Zampini } 1128b7ab4a40SStefano Zampini 112908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1130b7ab4a40SStefano Zampini if (compute_range) { 1131d16cbb6bSStefano Zampini 1132d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 113308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 11349d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 113508122e43SStefano Zampini #else 11369d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 113708122e43SStefano Zampini #endif 1138b7ab4a40SStefano Zampini } else if (!same_data) { 1139d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1140d16cbb6bSStefano Zampini B_IL = 1; 1141d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 11429d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1143d16cbb6bSStefano Zampini #else 11449d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1145d16cbb6bSStefano Zampini #endif 1146b7ab4a40SStefano Zampini } else { /* same_data is true, so get the adaptive function requested by the user */ 1147b7ab4a40SStefano Zampini PetscInt k; 1148b7ab4a40SStefano Zampini if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1149b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);CHKERRQ(ierr); 1150b7ab4a40SStefano Zampini ierr = PetscBLASIntCast(nmax,&B_neigs);CHKERRQ(ierr); 1151b7ab4a40SStefano Zampini nmin = nmax; 1152b7ab4a40SStefano Zampini ierr = PetscMemzero(eigv,subset_size*nmax*sizeof(PetscScalar));CHKERRQ(ierr); 1153b7ab4a40SStefano Zampini for (k=0;k<nmax;k++) { 1154b7ab4a40SStefano Zampini eigs[k] = 1./PETSC_SMALL; 1155b7ab4a40SStefano Zampini eigv[k*(subset_size+1)] = 1.0; 1156b7ab4a40SStefano Zampini } 1157d16cbb6bSStefano Zampini } 115808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 115908122e43SStefano Zampini if (B_ierr) { 11606c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 11616c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 11626c4ed002SBarry Smith else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 116308122e43SStefano Zampini } 116408122e43SStefano Zampini 116508122e43SStefano Zampini if (B_neigs > nmax) { 1166fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1167fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1168fd14bc51SStefano Zampini } 11699d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) eigs_start = B_neigs -nmax; 117008122e43SStefano Zampini B_neigs = nmax; 117108122e43SStefano Zampini } 117208122e43SStefano Zampini 11739552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11749552c7c7SStefano Zampini if (B_neigs < nmin_s) { 117508122e43SStefano Zampini PetscBLASInt B_neigs2; 117608122e43SStefano Zampini 11779d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1178f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 11799d54b7f4SStefano Zampini B_IU = B_N - B_neigs; 11809d54b7f4SStefano Zampini } else { 11819d54b7f4SStefano Zampini B_IL = B_neigs + 1; 11829d54b7f4SStefano Zampini B_IU = nmin_s; 11839d54b7f4SStefano Zampini } 1184fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1185fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 1186fd14bc51SStefano Zampini } 11879ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 118808122e43SStefano Zampini PetscInt j; 118908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 119008122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 119108122e43SStefano Zampini } 119208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 119308122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 119408122e43SStefano Zampini } 119508122e43SStefano Zampini } else { 119608122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119708122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119808122e43SStefano Zampini } 119908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 120008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 12019d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 120208122e43SStefano Zampini #else 12039d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 120408122e43SStefano Zampini #endif 120508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 120608122e43SStefano Zampini B_neigs += B_neigs2; 120708122e43SStefano Zampini } 120808122e43SStefano Zampini if (B_ierr) { 12096c4ed002SBarry Smith if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 12106c4ed002SBarry Smith else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 12116c4ed002SBarry Smith else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 121208122e43SStefano Zampini } 1213fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1214ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 121508122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 121608122e43SStefano Zampini if (eigs[j] == 0.0) { 1217ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 121808122e43SStefano Zampini } else { 12199d54b7f4SStefano Zampini if (pcbddc->use_deluxe_scaling) { 1220ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 12219d54b7f4SStefano Zampini } else { 12229d54b7f4SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);CHKERRQ(ierr); 12239d54b7f4SStefano Zampini } 1224fd14bc51SStefano Zampini } 122508122e43SStefano Zampini } 122608122e43SStefano Zampini } 122708122e43SStefano Zampini } else { 122808122e43SStefano Zampini /* TODO */ 122908122e43SStefano Zampini } 1230aff50787SStefano Zampini } 12316c3e6151SStefano Zampini /* change the basis back to the original one */ 12326c3e6151SStefano Zampini if (sub_schurs->change) { 123372b8c272SStefano Zampini Mat change,phi,phit; 12346c3e6151SStefano Zampini 12356c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12366c3e6151SStefano Zampini PetscInt ii; 12376c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12386c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12396c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12406c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12416c3e6151SStefano Zampini } 12426c3e6151SStefano Zampini } 12436c3e6151SStefano Zampini } 124472b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12456c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 124672b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12476c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12486c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12496c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12506c3e6151SStefano Zampini } 12518bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12528bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12539162d606SStefano Zampini if (B_neigs) { 12549162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1255fd14bc51SStefano Zampini 1256fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12579552c7c7SStefano Zampini PetscInt ii; 12589552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1259ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12609552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1261ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1262ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1263ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1264ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1265ac47001eSStefano Zampini #else 1266ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr); 1267ac47001eSStefano Zampini #endif 12689552c7c7SStefano Zampini } 12699552c7c7SStefano Zampini } 1270fd14bc51SStefano Zampini } 12719162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12729162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12739162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12749162d606SStefano Zampini cum++; 127508122e43SStefano Zampini } 127608122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 127708122e43SStefano Zampini /* shift for next computation */ 127808122e43SStefano Zampini cumarray += subset_size*subset_size; 127908122e43SStefano Zampini } 1280fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1281fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1282fd14bc51SStefano Zampini } 128308122e43SStefano Zampini 128408122e43SStefano Zampini if (mss) { 128508122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 128608122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1287f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1288f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1289f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 129008122e43SStefano Zampini } 1291f6f667cfSStefano Zampini if (allocated_S_St) { 1292f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1293f6f667cfSStefano Zampini } 1294f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 129508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 129608122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 129708122e43SStefano Zampini #endif 129808122e43SStefano Zampini if (pcbddc->dbg_flag) { 12991b968477SStefano Zampini PetscInt maxneigs_r; 1300b2566f29SBarry Smith ierr = MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 13019b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 130208122e43SStefano Zampini } 130308122e43SStefano Zampini PetscFunctionReturn(0); 130408122e43SStefano Zampini } 1305b1b3d7a2SStefano Zampini 1306674ae819SStefano Zampini #undef __FUNCT__ 1307c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1308c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1309c8587f34SStefano Zampini { 1310c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13118629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1312c8587f34SStefano Zampini PetscErrorCode ierr; 1313c8587f34SStefano Zampini 1314c8587f34SStefano Zampini PetscFunctionBegin; 1315f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 13165e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1317c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1318c8587f34SStefano Zampini 1319684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 13200fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1321684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1322c8587f34SStefano Zampini 1323c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1324b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1325c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1326c8587f34SStefano Zampini } 1327c8587f34SStefano Zampini 13288629588bSStefano Zampini /* 13298629588bSStefano Zampini Setup local correction and local part of coarse basis. 13308629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13318629588bSStefano Zampini */ 133247f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13338629588bSStefano Zampini 13348629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13358629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13368629588bSStefano Zampini 13378629588bSStefano Zampini /* free */ 13388629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1339c8587f34SStefano Zampini PetscFunctionReturn(0); 1340c8587f34SStefano Zampini } 1341c8587f34SStefano Zampini 1342c8587f34SStefano Zampini #undef __FUNCT__ 1343674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1344674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1345674ae819SStefano Zampini { 1346674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1347674ae819SStefano Zampini PetscErrorCode ierr; 1348674ae819SStefano Zampini 1349674ae819SStefano Zampini PetscFunctionBegin; 1350674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1351674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 135230368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1353674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1354674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1355785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1356674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1357f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1358f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1359785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 136063602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 136163602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1362674ae819SStefano Zampini PetscFunctionReturn(0); 1363674ae819SStefano Zampini } 1364674ae819SStefano Zampini 1365674ae819SStefano Zampini #undef __FUNCT__ 1366674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1367674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1368674ae819SStefano Zampini { 1369674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13704f1b2e48SStefano Zampini PetscInt i; 1371674ae819SStefano Zampini PetscErrorCode ierr; 1372674ae819SStefano Zampini 1373674ae819SStefano Zampini PetscFunctionBegin; 1374b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1375674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 137616909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 13771dd7afcfSStefano Zampini ierr = VecDestroy(&pcbddc->work_change);CHKERRQ(ierr); 1378674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1379674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13804f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13814f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13824f1b2e48SStefano Zampini } 13834f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1384b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1385674ae819SStefano Zampini PetscFunctionReturn(0); 1386674ae819SStefano Zampini } 1387674ae819SStefano Zampini 1388674ae819SStefano Zampini #undef __FUNCT__ 1389674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1390674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1391674ae819SStefano Zampini { 1392674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1393674ae819SStefano Zampini PetscErrorCode ierr; 1394674ae819SStefano Zampini 1395674ae819SStefano Zampini PetscFunctionBegin; 1396674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 139758da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1398ca92afb2SStefano Zampini PetscScalar *array; 139906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 140006656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 140158da7f69SStefano Zampini } 1402674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1403674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 140415aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 140515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1406674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1407674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1408674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 140906656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1410674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1411674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14128ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1413674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1414674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1415674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1416f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1417f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1418f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1419f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1420727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 14210e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1422f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 142370cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 14246e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 142581d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14260369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14271dd7afcfSStefano Zampini ierr = PCBDDCBenignShellMat(pc,PETSC_TRUE);CHKERRQ(ierr); 14284f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14298b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1430ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1431ca92afb2SStefano Zampini PetscInt i; 1432ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1433ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1434ca92afb2SStefano Zampini } 1435ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1436ca92afb2SStefano Zampini } 14374f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1438674ae819SStefano Zampini PetscFunctionReturn(0); 1439674ae819SStefano Zampini } 1440674ae819SStefano Zampini 1441674ae819SStefano Zampini #undef __FUNCT__ 1442f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1443f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14446bfb1811SStefano Zampini { 14456bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14466bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14476bfb1811SStefano Zampini VecType impVecType; 14484f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14496bfb1811SStefano Zampini PetscErrorCode ierr; 14506bfb1811SStefano Zampini 14516bfb1811SStefano Zampini PetscFunctionBegin; 14526c4ed002SBarry Smith if (!pcbddc->ConstraintMatrix) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1453e7b262bdSStefano Zampini /* get sizes */ 14544f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1455b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14566bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1457e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1458e7b262bdSStefano Zampini /* R nodes */ 1459e7b262bdSStefano Zampini old_size = -1; 1460e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1461e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1462e7b262bdSStefano Zampini } 1463e7b262bdSStefano Zampini if (n_R != old_size) { 1464e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1465e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14666bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14676bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14686bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14696bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1470e7b262bdSStefano Zampini } 1471e7b262bdSStefano Zampini /* local primal dofs */ 1472e7b262bdSStefano Zampini old_size = -1; 1473e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1474e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1475e7b262bdSStefano Zampini } 1476e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1477e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 147883b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1479e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14806bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1481e7b262bdSStefano Zampini } 1482e7b262bdSStefano Zampini /* local explicit constraints */ 1483e7b262bdSStefano Zampini old_size = -1; 1484e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1485e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1486e7b262bdSStefano Zampini } 1487e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1488e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 148983b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 149083b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 149183b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 149283b7ccabSStefano Zampini } 14936bfb1811SStefano Zampini PetscFunctionReturn(0); 14946bfb1811SStefano Zampini } 14956bfb1811SStefano Zampini 14966bfb1811SStefano Zampini #undef __FUNCT__ 149747f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 149847f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 149988ebb749SStefano Zampini { 150025084f0cSStefano Zampini PetscErrorCode ierr; 150125084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 150288ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 150388ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1504d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 150525084f0cSStefano Zampini /* submatrices of local problem */ 150680677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 150706656605SStefano Zampini /* submatrices of local coarse problem */ 150806656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 150925084f0cSStefano Zampini /* working matrices */ 151006656605SStefano Zampini Mat C_CR; 151125084f0cSStefano Zampini /* additional working stuff */ 151206656605SStefano Zampini PC pc_R; 15134f1b2e48SStefano Zampini Mat F; 15145cbda25cSStefano Zampini Vec dummy_vec; 1515a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 151625084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 151706656605SStefano Zampini PetscScalar *work; 151806656605SStefano Zampini PetscInt *idx_V_B; 1519ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 152006656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1521ffd830a3SStefano Zampini 152225084f0cSStefano Zampini /* some shortcuts to scalars */ 152306656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 152488ebb749SStefano Zampini 152588ebb749SStefano Zampini PetscFunctionBegin; 1526ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1527ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1528ffd830a3SStefano Zampini } 1529ffd830a3SStefano Zampini 1530ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1531b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15324f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1533b371cd4fSStefano Zampini n_B = pcis->n_B; 1534b371cd4fSStefano Zampini n_D = pcis->n - n_B; 153588ebb749SStefano Zampini n_R = pcis->n - n_vertices; 153688ebb749SStefano Zampini 153788ebb749SStefano Zampini /* vertices in boundary numbering */ 1538785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15390e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 15406c4ed002SBarry 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); 154188ebb749SStefano Zampini 154206656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1543019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 154406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 154506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 154606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 154706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 154806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 154906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 155106656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 155206656605SStefano Zampini 155306656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 155406656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 155506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 155606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 155706656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1558ffd830a3SStefano Zampini lda_rhs = n_R; 1559a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 156006656605SStefano Zampini if (isLU || isILU || isCHOL) { 156106656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1562df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1563df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1564d62866d3SStefano Zampini MatFactorType type; 1565d62866d3SStefano Zampini 1566df4d28bfSStefano Zampini F = reuse_solver->F; 15676816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1568d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1569ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1570df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 157106656605SStefano Zampini } else { 157206656605SStefano Zampini F = NULL; 157306656605SStefano Zampini } 157406656605SStefano Zampini 1575ffd830a3SStefano Zampini /* allocate workspace */ 1576ffd830a3SStefano Zampini n = 0; 1577ffd830a3SStefano Zampini if (n_constraints) { 1578ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1579ffd830a3SStefano Zampini } 1580ffd830a3SStefano Zampini if (n_vertices) { 1581ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1582ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1583ffd830a3SStefano Zampini } 1584ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1585ffd830a3SStefano Zampini 15865cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 15875cbda25cSStefano Zampini dummy_vec = NULL; 15885cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 15895cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 15905cbda25cSStefano Zampini } 15915cbda25cSStefano Zampini 159288ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 159388ebb749SStefano Zampini if (n_constraints) { 159472b8c272SStefano Zampini Mat M1,M2,M3,C_B; 159506656605SStefano Zampini IS is_aux; 159680677318SStefano Zampini PetscScalar *array,*array2; 159706656605SStefano Zampini 1598f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 159980677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 160088ebb749SStefano Zampini 160125084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 160225084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 16038ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 160472b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 160588ebb749SStefano Zampini 160680677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 160780677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1608ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 160988ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 161006656605SStefano Zampini const PetscScalar *row_cmat_values; 161106656605SStefano Zampini const PetscInt *row_cmat_indices; 161206656605SStefano Zampini PetscInt size_of_constraint,j; 161388ebb749SStefano Zampini 161406656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161506656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1616ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 161706656605SStefano Zampini } 161806656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161906656605SStefano Zampini } 1620ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 162106656605SStefano Zampini if (F) { 162206656605SStefano Zampini Mat B; 162306656605SStefano Zampini 1624ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1625a3df083aSStefano Zampini if (need_benign_correction) { 1626df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1627a3df083aSStefano Zampini 162872b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 162972b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1630a3df083aSStefano Zampini } 163180677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1632a3df083aSStefano Zampini if (need_benign_correction) { 1633a3df083aSStefano Zampini PetscScalar *marr; 1634df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1635a3df083aSStefano Zampini 1636a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 16375cbda25cSStefano Zampini if (lda_rhs != n_R) { 16385cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 16395cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 16405cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 16415cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 16425cbda25cSStefano Zampini } 16435cbda25cSStefano Zampini } else { 1644a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1645a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 16465cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1647a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1648a3df083aSStefano Zampini } 16495cbda25cSStefano Zampini } 1650a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1651a3df083aSStefano Zampini } 165206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 165306656605SStefano Zampini } else { 165480677318SStefano Zampini PetscScalar *marr; 165580677318SStefano Zampini 165680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 165706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1658ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1659ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 166006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 166106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 166206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 166306656605SStefano Zampini } 166480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 166506656605SStefano Zampini } 166680677318SStefano Zampini if (!pcbddc->switch_static) { 166780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 166880677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 166980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 167080677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1671ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 167280677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 167380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 167680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 167780677318SStefano Zampini } 167880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 167980677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 168072b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 168180677318SStefano Zampini } else { 1682ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1683ffd830a3SStefano Zampini IS dummy; 1684ffd830a3SStefano Zampini 1685ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 168672b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1687ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1688ffd830a3SStefano Zampini } else { 168980677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 169080677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1691ffd830a3SStefano Zampini } 169225084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 169380677318SStefano Zampini } 169480677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 169580677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 169680677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 169706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 169806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 169980677318SStefano Zampini if (isCHOL) { 170080677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 170180677318SStefano Zampini } else { 170225084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 170380677318SStefano Zampini } 170480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 170506656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 170625084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 170725084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 170825084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 170980677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 171072b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 171172b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 171206656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 171306656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1714f4ddd8eeSStefano Zampini } 1715fc227af8SStefano Zampini 1716fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 171788ebb749SStefano Zampini if (n_vertices) { 171806656605SStefano Zampini IS is_aux; 17193a50541eSStefano Zampini 1720df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 17216816873aSStefano Zampini IS tis; 17226816873aSStefano Zampini 17236816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17246816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17256816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17266816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17276816873aSStefano Zampini } else { 17283a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17296816873aSStefano Zampini } 17309577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17319577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 173204708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 173325084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 173488ebb749SStefano Zampini } 173588ebb749SStefano Zampini 173688ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1737f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 173806656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 173906656605SStefano Zampini if (pcbddc->coarse_phi_D) { 174006656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 174106656605SStefano Zampini } 1742f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 174306656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 174406656605SStefano Zampini PetscScalar *marray; 174506656605SStefano Zampini 174606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 174706656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1748f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1749f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1750f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1751f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1752f4ddd8eeSStefano Zampini } 1753f4ddd8eeSStefano Zampini } 175406656605SStefano Zampini 1755f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 175606656605SStefano Zampini PetscScalar *marray; 175788ebb749SStefano Zampini 175806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17598eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 176006656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 176188ebb749SStefano Zampini } 17623301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 176306656605SStefano Zampini n *= 2; 176488ebb749SStefano Zampini } 176506656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 176606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 176706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17688eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 176906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 177006656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 177188ebb749SStefano Zampini } 17723301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 177306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17748eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 177506656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 177606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 177788ebb749SStefano Zampini } 177888ebb749SStefano Zampini } else { 1779c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1780c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17811b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1782c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1783c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1784c0553b1fSStefano Zampini } 178588ebb749SStefano Zampini } 178606656605SStefano Zampini } 1787019a44ceSStefano Zampini 178806656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17894f1b2e48SStefano Zampini p0_lidx_I = NULL; 17904f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1791d12edf2fSStefano Zampini const PetscInt *idxs; 1792d12edf2fSStefano Zampini 1793d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17944f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17954f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17964f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17974f1b2e48SStefano Zampini } 1798d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1799d12edf2fSStefano Zampini } 1800d16cbb6bSStefano Zampini 180106656605SStefano Zampini /* vertices */ 180206656605SStefano Zampini if (n_vertices) { 180316f15bc4SStefano Zampini 1804af25d912SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);CHKERRQ(ierr); 180504708bb6SStefano Zampini 180616f15bc4SStefano Zampini if (n_R) { 180714393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 180806656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 180916f15bc4SStefano Zampini PetscScalar *x,*y; 181004708bb6SStefano Zampini PetscBool isseqaij; 181106656605SStefano Zampini 181221eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 181314393ed6SStefano Zampini if (need_benign_correction) { 181414393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 181514393ed6SStefano Zampini IS is_p0; 181614393ed6SStefano Zampini PetscInt *idxs_p0,n; 181714393ed6SStefano Zampini 181814393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 181914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 182014393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 1821af25d912SStefano 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); 182214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 182314393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 182414393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 182514393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 182614393ed6SStefano Zampini } 182714393ed6SStefano Zampini 1828ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1829af25d912SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1830ffd830a3SStefano Zampini } else { 1831ca92afb2SStefano Zampini PetscScalar *av,*array; 1832ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1833ca92afb2SStefano Zampini PetscInt n; 1834ca92afb2SStefano Zampini PetscBool flg_row; 1835ffd830a3SStefano Zampini 1836ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1837ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 18389d54b7f4SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);CHKERRQ(ierr); 1839ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1840ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1841ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1842ca92afb2SStefano Zampini PetscInt j; 1843ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1844ffd830a3SStefano Zampini } 1845ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1846ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1847ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1848ffd830a3SStefano Zampini } 1849ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1850a3df083aSStefano Zampini if (need_benign_correction) { 1851df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1852a3df083aSStefano Zampini PetscScalar *marr; 1853a3df083aSStefano Zampini 1854a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 185514393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 185614393ed6SStefano Zampini 185714393ed6SStefano Zampini | 0 0 0 | (V) 185814393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 185914393ed6SStefano Zampini | 0 0 -1 | (p0) 186014393ed6SStefano Zampini 186114393ed6SStefano Zampini */ 1862df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 186314393ed6SStefano Zampini const PetscScalar *vals; 186414393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 186514393ed6SStefano Zampini PetscInt n,j,nz; 186614393ed6SStefano Zampini 1867df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1868df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 187014393ed6SStefano Zampini for (j=0;j<n;j++) { 187114393ed6SStefano Zampini PetscScalar val = vals[j]; 187214393ed6SStefano Zampini PetscInt k,col = idxs[j]; 187314393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 187414393ed6SStefano Zampini } 187514393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1876df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 187714393ed6SStefano Zampini } 187872b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 187972b8c272SStefano Zampini } 188072b8c272SStefano Zampini if (F) { 188114393ed6SStefano Zampini /* need to correct the rhs */ 188272b8c272SStefano Zampini if (need_benign_correction) { 188372b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 188472b8c272SStefano Zampini PetscScalar *marr; 188572b8c272SStefano Zampini 188672b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 18875cbda25cSStefano Zampini if (lda_rhs != n_R) { 18885cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 18895cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 18905cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 18915cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 18925cbda25cSStefano Zampini } 18935cbda25cSStefano Zampini } else { 1894a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1895a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 18965cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1897a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1898a3df083aSStefano Zampini } 18995cbda25cSStefano Zampini } 1900a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1901a3df083aSStefano Zampini } 190206656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 190314393ed6SStefano Zampini /* need to correct the solution */ 1904a3df083aSStefano Zampini if (need_benign_correction) { 1905df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1906a3df083aSStefano Zampini PetscScalar *marr; 1907a3df083aSStefano Zampini 1908a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 19095cbda25cSStefano Zampini if (lda_rhs != n_R) { 19105cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 19115cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 19125cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 19135cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 19145cbda25cSStefano Zampini } 19155cbda25cSStefano Zampini } else { 1916a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1917a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 19185cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1919a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1920a3df083aSStefano Zampini } 19215cbda25cSStefano Zampini } 1922a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1923a3df083aSStefano Zampini } 192406656605SStefano Zampini } else { 192506656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 192606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1927ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1928ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 192906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 193006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 193106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 193206656605SStefano Zampini } 193306656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 193406656605SStefano Zampini } 193580677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1936ffd830a3SStefano Zampini /* S_VV and S_CV */ 193706656605SStefano Zampini if (n_constraints) { 193806656605SStefano Zampini Mat B; 193980677318SStefano Zampini 1940ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 194180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1942ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1943ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 194480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 194780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 194880677318SStefano Zampini } 1949ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 195080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 195180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1952ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 195380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 195406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1955ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1956ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 195706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195806656605SStefano Zampini } 195904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 196004708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 1961511c6705SHong Zhang ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 196204708bb6SStefano Zampini } 1963ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1964ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1965ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1966ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1967ffd830a3SStefano Zampini } 196806656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 196914393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 197014393ed6SStefano Zampini if (need_benign_correction) { 1971df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 197214393ed6SStefano Zampini PetscScalar *marr,*sums; 197314393ed6SStefano Zampini 197414393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 197514393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1976df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 197714393ed6SStefano Zampini const PetscScalar *vals; 197814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 197914393ed6SStefano Zampini PetscInt n,j,nz; 198014393ed6SStefano Zampini 1981df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1982df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 198314393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 198414393ed6SStefano Zampini PetscInt k; 198514393ed6SStefano Zampini sums[j] = 0.; 198614393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 198714393ed6SStefano Zampini } 198814393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 198914393ed6SStefano Zampini for (j=0;j<n;j++) { 199014393ed6SStefano Zampini PetscScalar val = vals[j]; 199114393ed6SStefano Zampini PetscInt k; 199214393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 199314393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 199414393ed6SStefano Zampini } 199514393ed6SStefano Zampini } 199614393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1997df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 199814393ed6SStefano Zampini } 199914393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 200014393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 200114393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 200214393ed6SStefano Zampini } 200380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 200406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 200506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 200606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 200706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 200806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 200906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 201006656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2011d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2012019a44ceSStefano Zampini } else { 2013d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2014d16cbb6bSStefano Zampini } 201521eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2016d16cbb6bSStefano Zampini 201706656605SStefano Zampini /* coarse basis functions */ 201806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 201916f15bc4SStefano Zampini PetscScalar *y; 202016f15bc4SStefano Zampini 2021ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 202206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 202406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 202706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 202906656605SStefano Zampini 203006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20314f1b2e48SStefano Zampini PetscInt j; 20324f1b2e48SStefano Zampini 203306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 203506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20384f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 203906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 204006656605SStefano Zampini } 204106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 204206656605SStefano Zampini } 204304708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 204404708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 204506656605SStefano Zampini } 20465cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 204706656605SStefano Zampini 204806656605SStefano Zampini if (n_constraints) { 204906656605SStefano Zampini Mat B; 205006656605SStefano Zampini 2051ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 205206656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 205406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205506656605SStefano Zampini if (n_vertices) { 205680677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 205780677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 205880677318SStefano Zampini } else { 205980677318SStefano Zampini Mat S_VCt; 206080677318SStefano Zampini 2061ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2062ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206372b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2064ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2065ffd830a3SStefano Zampini } 206680677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 206780677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 206880677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 206980677318SStefano Zampini } 207006656605SStefano Zampini } 207106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 207206656605SStefano Zampini /* coarse basis functions */ 207306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 207406656605SStefano Zampini PetscScalar *y; 207506656605SStefano Zampini 2076ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 207706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 207906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 208206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 208306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20844f1b2e48SStefano Zampini PetscInt j; 20854f1b2e48SStefano Zampini 208606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 208706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 208806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 209006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20914f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 209206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 209306656605SStefano Zampini } 209406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209506656605SStefano Zampini } 209606656605SStefano Zampini } 209780677318SStefano Zampini if (n_constraints) { 209880677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 209980677318SStefano Zampini } 21004f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 210172b8c272SStefano Zampini 210272b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 210372b8c272SStefano Zampini if (pcbddc->benign_n) { 210472b8c272SStefano Zampini Mat B0_B,B0_BPHI; 210572b8c272SStefano Zampini IS is_dummy; 210672b8c272SStefano Zampini PetscScalar *data; 210772b8c272SStefano Zampini PetscInt j; 210872b8c272SStefano Zampini 210972b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 211072b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 211172b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 211272b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 211386c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 211472b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 211572b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 211672b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 211772b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 211872b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 211972b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 212072b8c272SStefano Zampini } 212172b8c272SStefano Zampini } 212272b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 212372b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 212472b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 212572b8c272SStefano Zampini } 2126019a44ceSStefano Zampini 212706656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 21283301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2129ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2130ffd830a3SStefano Zampini PetscScalar *marray; 213106656605SStefano Zampini 213206656605SStefano Zampini if (n_constraints) { 2133ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 213406656605SStefano Zampini 2135af25d912SStefano Zampini ierr = MatTranspose(C_CR,MAT_INPLACE_MATRIX,&C_CRT);CHKERRQ(ierr); 213606656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2137ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 213816f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 213906656605SStefano Zampini if (n_vertices) { 2140ffd830a3SStefano Zampini Mat S_VCT; 214106656605SStefano Zampini 214206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2143ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 214416f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 214506656605SStefano Zampini } 2146ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 214706656605SStefano Zampini } 214816f15bc4SStefano Zampini if (n_vertices && n_R) { 2149ffd830a3SStefano Zampini PetscScalar *av,*marray; 2150ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2151ffd830a3SStefano Zampini PetscInt n; 2152ffd830a3SStefano Zampini PetscBool flg_row; 215306656605SStefano Zampini 2154ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2155af25d912SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);CHKERRQ(ierr); 2156ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2157ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2158ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2159ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2160ffd830a3SStefano Zampini PetscInt j; 2161ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2162ffd830a3SStefano Zampini } 2163ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2164ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2165ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 216606656605SStefano Zampini } 216706656605SStefano Zampini 2168ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2169ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2170ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2171ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2172ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 217306656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 217406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 217606656605SStefano Zampini } 2177ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2178ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2179ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2180ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2181ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2182ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2183ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2184ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 218506656605SStefano Zampini } 2186ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 218706656605SStefano Zampini /* coarse basis functions */ 218806656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 218906656605SStefano Zampini PetscScalar *y; 219006656605SStefano Zampini 2191ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 219206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 219406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219606656605SStefano Zampini if (i<n_vertices) { 219706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 219806656605SStefano Zampini } 219906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 220006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 220106656605SStefano Zampini 220206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 220306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 220506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 220806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220906656605SStefano Zampini } 221006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 221106656605SStefano Zampini } 2212ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2213ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 221406656605SStefano Zampini } 2215d62866d3SStefano Zampini /* free memory */ 221688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 221706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 221806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 221906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 222006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2221d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2222d62866d3SStefano Zampini if (n_vertices) { 2223d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2224d62866d3SStefano Zampini } 2225d62866d3SStefano Zampini if (n_constraints) { 2226d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2227d62866d3SStefano Zampini } 222888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 222988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 223088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2231d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 223288ebb749SStefano Zampini Mat coarse_sub_mat; 223325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 223488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 223588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 223688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22378bec7fa6SStefano Zampini Mat C_B,CPHI; 22388bec7fa6SStefano Zampini IS is_dummy; 22398bec7fa6SStefano Zampini Vec mones; 224088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 224188ebb749SStefano Zampini PetscReal real_value; 224288ebb749SStefano Zampini 2243a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2244a3df083aSStefano Zampini Mat A; 2245a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2246a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2247a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2248a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2249a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2250a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2251a3df083aSStefano Zampini } else { 225288ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 225388ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 225488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 225588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2256a3df083aSStefano Zampini } 225788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 225888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2259ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 226188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 226288ebb749SStefano Zampini } 226388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 226488ebb749SStefano Zampini 226525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22663301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 226725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2268ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 227188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 227788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 228088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228188ebb749SStefano Zampini } else { 228288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 228388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 228488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 228688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 228988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 229088ebb749SStefano Zampini } 229188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 2294511c6705SHong Zhang ierr = MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);CHKERRQ(ierr); 22954f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2296fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2297d12edf2fSStefano Zampini PetscScalar *data,*data2; 22984f1b2e48SStefano Zampini PetscInt j; 2299d12edf2fSStefano Zampini 23004f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2301fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2302d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 230386c38910SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2304d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2305d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 23064f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 23074f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2308d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 23094f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 23104f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 23114f1b2e48SStefano Zampini } 2312d12edf2fSStefano Zampini } 2313d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2314d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2315d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2316d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2317d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2318d12edf2fSStefano Zampini } 2319d12edf2fSStefano Zampini #if 0 2320d12edf2fSStefano Zampini { 2321d12edf2fSStefano Zampini PetscViewer viewer; 2322d12edf2fSStefano Zampini char filename[256]; 2323ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2324d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2325d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2326ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2327ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2328ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2329d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 233072b8c272SStefano Zampini if (save_change) { 233172b8c272SStefano Zampini Mat phi_B; 233272b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 233372b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 233472b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 233572b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 233672b8c272SStefano Zampini } else { 2337ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2338ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 233972b8c272SStefano Zampini } 2340ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2341ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2342ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2343ffd830a3SStefano Zampini } 2344ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2345ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2346ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2347ffd830a3SStefano Zampini } 234872b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2349ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2350ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2351ffd830a3SStefano Zampini } 2352d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2353d12edf2fSStefano Zampini } 2354d12edf2fSStefano Zampini #endif 235581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23568bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23571575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 235806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23598bec7fa6SStefano Zampini 23608bec7fa6SStefano Zampini /* check constraints */ 2361a00504b5SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2362a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 23634f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23648bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2365a00504b5SStefano Zampini } else { 2366a00504b5SStefano Zampini PetscScalar *data; 2367a00504b5SStefano Zampini Mat tmat; 2368a00504b5SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2369a00504b5SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);CHKERRQ(ierr); 2370a00504b5SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);CHKERRQ(ierr); 2371a00504b5SStefano Zampini ierr = MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2372a00504b5SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 2373a00504b5SStefano Zampini } 23748bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23758bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23768bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23778bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2378bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2379ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2380bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2381bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2382bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2383bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2384bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 238588ebb749SStefano Zampini } 23868bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23878bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23888bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23898bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 239025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 239188ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 239288ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 239388ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 239488ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 239588ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 239688ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 239788ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 239888ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 239988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 240088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2401ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 240288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 240388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 240488ebb749SStefano Zampini } 240588ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 240688ebb749SStefano Zampini } 24078629588bSStefano Zampini /* get back data */ 24088629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 240988ebb749SStefano Zampini PetscFunctionReturn(0); 241088ebb749SStefano Zampini } 241188ebb749SStefano Zampini 241288ebb749SStefano Zampini #undef __FUNCT__ 2413d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2414d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2415aa0d41d4SStefano Zampini { 2416d65f70fdSStefano Zampini Mat *work_mat; 2417d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2418d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2419d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2420aa0d41d4SStefano Zampini PetscErrorCode ierr; 2421aa0d41d4SStefano Zampini 2422aa0d41d4SStefano Zampini PetscFunctionBegin; 2423d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2424d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2425d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2426d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2427aa0d41d4SStefano Zampini 2428d65f70fdSStefano Zampini if (!rsorted) { 2429906d46d4SStefano Zampini const PetscInt *idxs; 2430906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2431aa0d41d4SStefano Zampini 2432d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2433d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2434d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2435d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2436aa0d41d4SStefano Zampini } 2437d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2438d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2439d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2440d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2441aa0d41d4SStefano Zampini } 2442d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2443d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2444d65f70fdSStefano Zampini } else { 2445d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2446d65f70fdSStefano Zampini isrow_s = isrow; 2447aa0d41d4SStefano Zampini } 2448906d46d4SStefano Zampini 2449d65f70fdSStefano Zampini if (!csorted) { 2450d65f70fdSStefano Zampini if (isrow == iscol) { 2451d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2452d65f70fdSStefano Zampini iscol_s = isrow_s; 2453d65f70fdSStefano Zampini } else { 2454d65f70fdSStefano Zampini const PetscInt *idxs; 2455d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2456906d46d4SStefano Zampini 2457d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2458d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2459d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2460d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2461d65f70fdSStefano Zampini } 2462d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2463d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2464d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2465d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2466d65f70fdSStefano Zampini } 2467d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2468d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2469d65f70fdSStefano Zampini } 2470d65f70fdSStefano Zampini } else { 2471d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2472d65f70fdSStefano Zampini iscol_s = iscol; 2473d65f70fdSStefano Zampini } 2474d65f70fdSStefano Zampini 2475d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2476d65f70fdSStefano Zampini 2477d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2478906d46d4SStefano Zampini Mat new_mat; 2479d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2480906d46d4SStefano Zampini 2481d65f70fdSStefano Zampini if (!rsorted) { 2482d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2483d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2484d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2485d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2486906d46d4SStefano Zampini } 2487d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2488d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2489d65f70fdSStefano Zampini } else { 2490d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2491906d46d4SStefano Zampini } 2492d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2493d65f70fdSStefano Zampini 2494d65f70fdSStefano Zampini if (!csorted) { 2495d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2496d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2497d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2498d65f70fdSStefano Zampini } else { 2499d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2500d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2501d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2502d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2503d65f70fdSStefano Zampini } 2504d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2505d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2506d65f70fdSStefano Zampini } 2507d65f70fdSStefano Zampini } else { 2508d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2509d65f70fdSStefano Zampini } 2510d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2511d65f70fdSStefano Zampini 2512d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2513d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2514d65f70fdSStefano Zampini work_mat[0] = new_mat; 2515d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2516d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2517d65f70fdSStefano Zampini } 2518d65f70fdSStefano Zampini 2519d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2520d65f70fdSStefano Zampini *B = work_mat[0]; 2521d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2522d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2523d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2524d65f70fdSStefano Zampini PetscFunctionReturn(0); 2525d65f70fdSStefano Zampini } 2526d65f70fdSStefano Zampini 2527d65f70fdSStefano Zampini #undef __FUNCT__ 25285e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 25295e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2530aa0d41d4SStefano Zampini { 2531aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 25325e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2533d65f70fdSStefano Zampini Mat new_mat; 25345e8657edSStefano Zampini IS is_local,is_global; 2535d65f70fdSStefano Zampini PetscInt local_size; 2536d65f70fdSStefano Zampini PetscBool isseqaij; 2537aa0d41d4SStefano Zampini PetscErrorCode ierr; 2538aa0d41d4SStefano Zampini 2539aa0d41d4SStefano Zampini PetscFunctionBegin; 2540aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25415e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25425e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2543b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2544aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2545d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2546aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2547906d46d4SStefano Zampini 2548906d46d4SStefano Zampini /* check */ 2549906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2550906d46d4SStefano Zampini Vec x,x_change; 2551906d46d4SStefano Zampini PetscReal error; 2552906d46d4SStefano Zampini 25535e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2554906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25555e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2556e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2557e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2558d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2559e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2560e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2561906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2562906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2563906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2564906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2565906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2566906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2567906d46d4SStefano Zampini } 2568906d46d4SStefano Zampini 256922d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25709b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 257122d5777bSStefano Zampini if (isseqaij) { 2572a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2573a00504b5SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2574aa0d41d4SStefano Zampini } else { 2575a00504b5SStefano Zampini Mat work_mat; 25761cf9b237SStefano Zampini 2577a00504b5SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 2578aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2579a00504b5SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);CHKERRQ(ierr); 2580aa0d41d4SStefano Zampini } 25813301b35fSStefano Zampini if (matis->A->symmetric_set) { 25823301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2583e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25843301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2585e496cd5dSStefano Zampini #endif 25863301b35fSStefano Zampini } 2587d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2588aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2589aa0d41d4SStefano Zampini } 2590aa0d41d4SStefano Zampini 2591aa0d41d4SStefano Zampini #undef __FUNCT__ 2592a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25938ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2594a64d13efSStefano Zampini { 2595a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2596a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2597d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 259853892102SStefano Zampini PetscInt *idx_R_local=NULL; 25993a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 26003a50541eSStefano Zampini PetscInt vbs,bs; 26016816873aSStefano Zampini PetscBT bitmask=NULL; 2602a64d13efSStefano Zampini PetscErrorCode ierr; 2603a64d13efSStefano Zampini 2604a64d13efSStefano Zampini PetscFunctionBegin; 2605b23d619eSStefano Zampini /* 2606b23d619eSStefano Zampini No need to setup local scatters if 2607b23d619eSStefano Zampini - primal space is unchanged 2608b23d619eSStefano Zampini AND 2609b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2610b23d619eSStefano Zampini AND 2611b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2612b23d619eSStefano Zampini */ 2613b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2614f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2615f4ddd8eeSStefano Zampini } 2616f4ddd8eeSStefano Zampini /* destroy old objects */ 2617f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2618f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2619f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2620a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2621b371cd4fSStefano Zampini n_B = pcis->n_B; 2622b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2623b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 26243a50541eSStefano Zampini 2625a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 26266816873aSStefano Zampini 262753892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2628df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2629854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2630a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2631a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 26320e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2633a64d13efSStefano Zampini } 2634a64d13efSStefano Zampini 2635a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26364641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26376816873aSStefano Zampini idx_R_local[n_R++] = i; 2638a64d13efSStefano Zampini } 2639a64d13efSStefano Zampini } 2640df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2641df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26426816873aSStefano Zampini 2643df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2644df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26456816873aSStefano Zampini } 26463a50541eSStefano Zampini 26473a50541eSStefano Zampini /* Block code */ 26483a50541eSStefano Zampini vbs = 1; 26493a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26503a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26513a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26523a50541eSStefano Zampini PetscInt *vary; 2653df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2654785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26553a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2656d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2657d3df7717SStefano 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 */ 26580e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2659d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26603a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26613a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26623a50541eSStefano Zampini break; 26633a50541eSStefano Zampini } 26643a50541eSStefano Zampini } 2665d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2666d3df7717SStefano Zampini } else { 2667d3df7717SStefano Zampini /* Verify directly the R set */ 2668d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2669d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2670d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2671d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2672d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2673d3df7717SStefano Zampini break; 2674d3df7717SStefano Zampini } 2675d3df7717SStefano Zampini } 2676d3df7717SStefano Zampini } 2677d3df7717SStefano Zampini } 26783a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26793a50541eSStefano Zampini vbs = bs; 26803a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26813a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26823a50541eSStefano Zampini } 26833a50541eSStefano Zampini } 26843a50541eSStefano Zampini } 26853a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2686df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2687df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 268853892102SStefano Zampini 2689df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2690df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 269153892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2692df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 269353892102SStefano Zampini } else { 26943a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 269553892102SStefano Zampini } 2696a64d13efSStefano Zampini 2697a64d13efSStefano Zampini /* print some info if requested */ 2698a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2699a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2700a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 27011575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2702a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2703a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 27044f1b2e48SStefano 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); 2705a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2706a64d13efSStefano Zampini } 2707a64d13efSStefano Zampini 2708a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2709df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 27106816873aSStefano Zampini IS is_aux1,is_aux2; 27116816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 27126816873aSStefano Zampini 27133a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2714854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2715854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2716a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 27174641a718SStefano Zampini for (i=0; i<n_D; i++) { 27184641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 27194641a718SStefano Zampini } 2720a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2721a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27224641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 27234641a718SStefano Zampini aux_array1[j++] = i; 2724a64d13efSStefano Zampini } 2725a64d13efSStefano Zampini } 2726a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2727a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2728a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 27294641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 27304641a718SStefano Zampini aux_array2[j++] = i; 2731a64d13efSStefano Zampini } 2732a64d13efSStefano Zampini } 2733a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2734a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2735a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2736a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2737a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2738a64d13efSStefano Zampini 27398eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2740785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2741a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27424641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27434641a718SStefano Zampini aux_array1[j++] = i; 2744a64d13efSStefano Zampini } 2745a64d13efSStefano Zampini } 2746a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2747a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2748a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2749a64d13efSStefano Zampini } 27504641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27513a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2752d62866d3SStefano Zampini } else { 2753df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27546816873aSStefano Zampini IS tis; 27556816873aSStefano Zampini PetscInt schur_size; 27566816873aSStefano Zampini 2757df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27586816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2759df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27606816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27616816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27626816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27636816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27646816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2765d62866d3SStefano Zampini } 2766d62866d3SStefano Zampini } 2767a64d13efSStefano Zampini PetscFunctionReturn(0); 2768a64d13efSStefano Zampini } 2769a64d13efSStefano Zampini 2770304d26faSStefano Zampini 2771304d26faSStefano Zampini #undef __FUNCT__ 2772304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2773684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2774304d26faSStefano Zampini { 2775304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2776304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2777304d26faSStefano Zampini PC pc_temp; 2778304d26faSStefano Zampini Mat A_RR; 2779f4ddd8eeSStefano Zampini MatReuse reuse; 2780304d26faSStefano Zampini PetscScalar m_one = -1.0; 2781304d26faSStefano Zampini PetscReal value; 278204708bb6SStefano Zampini PetscInt n_D,n_R; 27839577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2784304d26faSStefano Zampini PetscErrorCode ierr; 2785e604994aSStefano Zampini /* prefixes stuff */ 2786312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2787e604994aSStefano Zampini size_t len; 2788304d26faSStefano Zampini 2789304d26faSStefano Zampini PetscFunctionBegin; 2790304d26faSStefano Zampini 2791e604994aSStefano Zampini /* compute prefixes */ 2792e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2793e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2794e604994aSStefano Zampini if (!pcbddc->current_level) { 2795e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2796e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2797e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2798e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2799e604994aSStefano Zampini } else { 2800e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2801312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2802e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2803e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2804312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2805312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 280634d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 280734d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2808e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2809e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2810e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2811e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2812e604994aSStefano Zampini } 2813e604994aSStefano Zampini 2814304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2815684f6988SStefano Zampini if (dirichlet) { 2816d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2817450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2818450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2819450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2820450f8f5eSStefano Zampini } 2821450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2822a3df083aSStefano Zampini Mat A_IIn; 2823a3df083aSStefano Zampini 2824a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2825a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2826a3df083aSStefano Zampini pcis->A_II = A_IIn; 2827a3df083aSStefano Zampini } 2828450f8f5eSStefano Zampini } 28293301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28303301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2831964fefecSStefano Zampini } 2832ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2833964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2834304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2835304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2836304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2837304d26faSStefano Zampini /* default */ 2838304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2839e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28409577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2841304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28429577ea80SStefano Zampini if (issbaij) { 28439577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28449577ea80SStefano Zampini } else { 2845304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28469577ea80SStefano Zampini } 2847304d26faSStefano Zampini /* Allow user's customization */ 2848304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2849304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2850304d26faSStefano Zampini } 2851d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2852df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2853df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2854d62866d3SStefano Zampini 2855df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2856d5574798SStefano Zampini } 2857304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2858304d26faSStefano Zampini if (!n_D) { 2859304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2860304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2861304d26faSStefano Zampini } 2862304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2863304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2864304d26faSStefano Zampini /* set ksp_D into pcis data */ 2865304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2866304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2867304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2868684f6988SStefano Zampini } 2869304d26faSStefano Zampini 2870304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2871684f6988SStefano Zampini A_RR = 0; 2872684f6988SStefano Zampini if (neumann) { 2873d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 287404708bb6SStefano Zampini PetscInt ibs,mbs; 287504708bb6SStefano Zampini PetscBool issbaij; 287604708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2877f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28788ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2879f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2880f4ddd8eeSStefano Zampini PetscInt nn_R; 288181d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2882f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2883f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2884f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2885f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2886f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2887f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2888f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2889727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2890f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2891f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2892f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2893f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2894f4ddd8eeSStefano Zampini } 2895f4ddd8eeSStefano Zampini } 2896f4ddd8eeSStefano Zampini /* last check */ 2897d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2898f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2899f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2900f4ddd8eeSStefano Zampini } 2901f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2902f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2903f4ddd8eeSStefano Zampini } 2904a00504b5SStefano Zampini /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection */ 2905af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2906af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 290704708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 290804708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 290904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2912af732b37SStefano Zampini } else { 2913511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 29146816873aSStefano Zampini } 291504708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 291604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 291904708bb6SStefano Zampini } else { 2920511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 292104708bb6SStefano Zampini } 292204708bb6SStefano Zampini } 2923a00504b5SStefano Zampini /* extract A_RR */ 2924a00504b5SStefano Zampini if (sub_schurs->reuse_solver) { 2925a00504b5SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2926a00504b5SStefano Zampini 2927a00504b5SStefano Zampini if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */ 292816e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2929a00504b5SStefano Zampini if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */ 293016e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 293116e386b8SStefano Zampini } else { 2932a00504b5SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);CHKERRQ(ierr); 2933a00504b5SStefano Zampini } 2934a00504b5SStefano Zampini } else { 2935a00504b5SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2936a00504b5SStefano Zampini ierr = PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);CHKERRQ(ierr); 2937a00504b5SStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2938a00504b5SStefano Zampini } 2939a00504b5SStefano Zampini } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */ 2940f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 294116e386b8SStefano Zampini } 29423301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29433301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29446816873aSStefano Zampini } 2945f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2946304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2947304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2948304d26faSStefano Zampini /* default */ 2949304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2950e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2951304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29529577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29539577ea80SStefano Zampini if (issbaij) { 29549577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29559577ea80SStefano Zampini } else { 2956304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29579577ea80SStefano Zampini } 2958304d26faSStefano Zampini /* Allow user's customization */ 2959304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2960304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2961304d26faSStefano Zampini } 2962304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2963304d26faSStefano Zampini if (!n_R) { 2964304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2965304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2966304d26faSStefano Zampini } 29675cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2968df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2969df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2970df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2971d62866d3SStefano Zampini 2972df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2973d62866d3SStefano Zampini } 2974304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2975304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2976684f6988SStefano Zampini } 2977304d26faSStefano Zampini 2978304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29790fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2980684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2981684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29821575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2983684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2984684f6988SStefano Zampini } 2985684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29860fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29870fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29880fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29890fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29900fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2991304d26faSStefano Zampini /* need to be adapted? */ 2992b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2993b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2994b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2995304d26faSStefano Zampini /* print info */ 2996304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2997e604994aSStefano 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); 2998304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2999304d26faSStefano Zampini } 3000b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 3001298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 3002304d26faSStefano Zampini } 3003684f6988SStefano Zampini } 3004684f6988SStefano Zampini if (neumann) { /* Neumann */ 30050fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 30060fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 30070fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 30080fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 30090fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 3010304d26faSStefano Zampini /* need to be adapted? */ 3011b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 3012b2566f29SBarry Smith ierr = MPIU_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3013304d26faSStefano Zampini /* print info */ 3014304d26faSStefano Zampini if (pcbddc->dbg_flag) { 3015e604994aSStefano 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); 3016304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3017304d26faSStefano Zampini } 3018b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 3019298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 3020304d26faSStefano Zampini } 30210fccc4e9SStefano Zampini } 3022684f6988SStefano Zampini } 30235cbda25cSStefano Zampini /* free Neumann problem's matrix */ 30245cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3025304d26faSStefano Zampini PetscFunctionReturn(0); 3026304d26faSStefano Zampini } 3027304d26faSStefano Zampini 3028304d26faSStefano Zampini #undef __FUNCT__ 3029ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 303080677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3031674ae819SStefano Zampini { 3032674ae819SStefano Zampini PetscErrorCode ierr; 3033674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3034be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3035674ae819SStefano Zampini 3036674ae819SStefano Zampini PetscFunctionBegin; 3037df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 303880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 303920c7b377SStefano Zampini } 304080677318SStefano Zampini if (!pcbddc->switch_static) { 304180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 304280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 304380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 304420c7b377SStefano Zampini } 3045df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 304680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304820c7b377SStefano Zampini } else { 3049df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3050be83ff47SStefano Zampini 3051df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3052df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305320c7b377SStefano Zampini } 3054be83ff47SStefano Zampini } else { 305580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 305980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 306080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 306180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 306280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3064674ae819SStefano Zampini } 3065674ae819SStefano Zampini } 306672b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 306780677318SStefano Zampini if (applytranspose) { 306880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 306980677318SStefano Zampini } else { 307080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 307180677318SStefano Zampini } 3072be83ff47SStefano Zampini } else { 3073df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3074be83ff47SStefano Zampini 3075be83ff47SStefano Zampini if (applytranspose) { 3076df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3077be83ff47SStefano Zampini } else { 3078df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3079be83ff47SStefano Zampini } 3080be83ff47SStefano Zampini } 308180677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 308280677318SStefano Zampini if (!pcbddc->switch_static) { 3083df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 308480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3086be83ff47SStefano Zampini } else { 3087df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3088be83ff47SStefano Zampini 3089df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3090df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3091be83ff47SStefano Zampini } 309280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 309380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 309480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 309580677318SStefano Zampini } 309680677318SStefano Zampini } else { 309780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 310280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 310380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 310480677318SStefano Zampini } 310580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3109674ae819SStefano Zampini } 3110674ae819SStefano Zampini PetscFunctionReturn(0); 3111674ae819SStefano Zampini } 3112674ae819SStefano Zampini 3113dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3114674ae819SStefano Zampini #undef __FUNCT__ 3115674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3116dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3117674ae819SStefano Zampini { 3118674ae819SStefano Zampini PetscErrorCode ierr; 3119674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3120674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3121674ae819SStefano Zampini const PetscScalar zero = 0.0; 3122674ae819SStefano Zampini 3123674ae819SStefano Zampini PetscFunctionBegin; 3124dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 31254fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3126dc359a40SStefano Zampini if (applytranspose) { 3127674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 31288eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3129dc359a40SStefano Zampini } else { 3130674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3131674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 313215aaf578SStefano Zampini } 31334fee134fSStefano Zampini } else { 31344fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 31354fee134fSStefano Zampini } 3136efc2fbd9SStefano Zampini 3137efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31384f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3139efc2fbd9SStefano Zampini PetscScalar *array; 31404f1b2e48SStefano Zampini PetscInt j; 3141efc2fbd9SStefano Zampini 3142efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31434f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3144efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3145efc2fbd9SStefano Zampini } 3146efc2fbd9SStefano Zampini 314712edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 314812edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 314912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 315112edc857SStefano Zampini 31529f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 315312edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 315412edc857SStefano Zampini if (pcbddc->coarse_ksp) { 315551694757SStefano Zampini Mat coarse_mat; 3156964fefecSStefano Zampini Vec rhs,sol; 315751694757SStefano Zampini MatNullSpace nullsp; 3158964fefecSStefano Zampini 3159964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3160964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 316151694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 316251694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 316351694757SStefano Zampini if (nullsp) { 316451694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 316551694757SStefano Zampini } 316612edc857SStefano Zampini if (applytranspose) { 3167964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 316812edc857SStefano Zampini } else { 3169964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 317012edc857SStefano Zampini } 317151694757SStefano Zampini if (nullsp) { 317251694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 317351694757SStefano Zampini } 317412edc857SStefano Zampini } 3175674ae819SStefano Zampini 3176674ae819SStefano Zampini /* Local solution on R nodes */ 31774fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 317880677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31799f00e9b4SStefano Zampini } 31809f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31819f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 318212edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3183674ae819SStefano Zampini 31844fee134fSStefano Zampini /* Sum contributions from the two levels */ 31854fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3186dc359a40SStefano Zampini if (applytranspose) { 3187dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3188dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3189dc359a40SStefano Zampini } else { 3190674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31918eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3192dc359a40SStefano Zampini } 3193efc2fbd9SStefano Zampini /* store p0 */ 31944f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3195efc2fbd9SStefano Zampini PetscScalar *array; 31964f1b2e48SStefano Zampini PetscInt j; 3197efc2fbd9SStefano Zampini 3198efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31994f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3200efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3201efc2fbd9SStefano Zampini } 32024fee134fSStefano Zampini } else { /* expand the coarse solution */ 32034fee134fSStefano Zampini if (applytranspose) { 32044fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32054fee134fSStefano Zampini } else { 32064fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 32074fee134fSStefano Zampini } 32084fee134fSStefano Zampini } 3209674ae819SStefano Zampini PetscFunctionReturn(0); 3210674ae819SStefano Zampini } 3211674ae819SStefano Zampini 3212674ae819SStefano Zampini #undef __FUNCT__ 3213674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 321412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3215674ae819SStefano Zampini { 3216674ae819SStefano Zampini PetscErrorCode ierr; 3217674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 321858da7f69SStefano Zampini PetscScalar *array; 321912edc857SStefano Zampini Vec from,to; 3220674ae819SStefano Zampini 3221674ae819SStefano Zampini PetscFunctionBegin; 322212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 322312edc857SStefano Zampini from = pcbddc->coarse_vec; 322412edc857SStefano Zampini to = pcbddc->vec1_P; 322512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 322612edc857SStefano Zampini Vec tvec; 322758da7f69SStefano Zampini 322858da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 322958da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 323012edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 323158da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 323258da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 323358da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 323412edc857SStefano Zampini } 323512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 323612edc857SStefano Zampini from = pcbddc->vec1_P; 323712edc857SStefano Zampini to = pcbddc->coarse_vec; 323812edc857SStefano Zampini } 323912edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3240674ae819SStefano Zampini PetscFunctionReturn(0); 3241674ae819SStefano Zampini } 3242674ae819SStefano Zampini 3243674ae819SStefano Zampini #undef __FUNCT__ 3244674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 324512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3246674ae819SStefano Zampini { 3247674ae819SStefano Zampini PetscErrorCode ierr; 3248674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 324958da7f69SStefano Zampini PetscScalar *array; 325012edc857SStefano Zampini Vec from,to; 3251674ae819SStefano Zampini 3252674ae819SStefano Zampini PetscFunctionBegin; 325312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 325412edc857SStefano Zampini from = pcbddc->coarse_vec; 325512edc857SStefano Zampini to = pcbddc->vec1_P; 325612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 325712edc857SStefano Zampini from = pcbddc->vec1_P; 325812edc857SStefano Zampini to = pcbddc->coarse_vec; 325912edc857SStefano Zampini } 326012edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 326112edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 326212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 326312edc857SStefano Zampini Vec tvec; 326458da7f69SStefano Zampini 326512edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 326658da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 326758da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 326858da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 326958da7f69SStefano Zampini } 327058da7f69SStefano Zampini } else { 327158da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 327258da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 327312edc857SStefano Zampini } 327412edc857SStefano Zampini } 3275674ae819SStefano Zampini PetscFunctionReturn(0); 3276674ae819SStefano Zampini } 3277674ae819SStefano Zampini 3278984c4197SStefano Zampini /* uncomment for testing purposes */ 3279984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3280674ae819SStefano Zampini #undef __FUNCT__ 3281674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3282674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3283674ae819SStefano Zampini { 3284674ae819SStefano Zampini PetscErrorCode ierr; 3285674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3286674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3287674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3288984c4197SStefano Zampini /* one and zero */ 3289984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3290984c4197SStefano Zampini /* space to store constraints and their local indices */ 32919162d606SStefano Zampini PetscScalar *constraints_data; 32929162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32939162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32949162d606SStefano Zampini PetscInt *constraints_n; 3295984c4197SStefano Zampini /* iterators */ 3296b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3297984c4197SStefano Zampini /* BLAS integers */ 3298e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3299e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3300c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3301727cdba6SStefano Zampini /* reuse */ 33020e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 33030e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3304984c4197SStefano Zampini /* change of basis */ 3305b3d85658SStefano Zampini PetscBool qr_needed; 33069162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3307984c4197SStefano Zampini /* auxiliary stuff */ 330864efe560SStefano Zampini PetscInt *nnz,*is_indices; 33098a0068c3SStefano Zampini PetscInt ncc; 3310984c4197SStefano Zampini /* some quantities */ 331145a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3312a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3313984c4197SStefano Zampini 3314674ae819SStefano Zampini PetscFunctionBegin; 33158e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 33168e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 33178e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 331816909a7fSStefano Zampini ierr = MatDestroy(&pcbddc->switch_static_change);CHKERRQ(ierr); 3319088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3320088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 33210e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 33220e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 33230e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33240e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33250e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3326088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3327cf5a6209SStefano Zampini 3328cf5a6209SStefano Zampini /* print some info */ 3329cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3330cf5a6209SStefano Zampini IS vertices; 3331cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3332cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3333cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3334cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 33351575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3336cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3337cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3338fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3339fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3340cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33411575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3342cf5a6209SStefano Zampini } 3343cf5a6209SStefano Zampini 3344cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33459162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3346cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3347cf5a6209SStefano Zampini const Vec *nearnullvecs; 3348cf5a6209SStefano Zampini Vec *localnearnullsp; 3349cf5a6209SStefano Zampini PetscScalar *array; 3350cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3351cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3352674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3353b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3354674ae819SStefano Zampini PetscScalar *work; 3355674ae819SStefano Zampini PetscReal *singular_vals; 3356674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3357674ae819SStefano Zampini PetscReal *rwork; 3358674ae819SStefano Zampini #endif 3359674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3360674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3361674ae819SStefano Zampini #else 3362964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3363964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3364674ae819SStefano Zampini #endif 3365674ae819SStefano Zampini 3366674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3367d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3368d06fc5fdSStefano Zampini /* free unneeded index sets */ 3369d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3370d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3371674ae819SStefano Zampini } 3372d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3373d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3374d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3375d06fc5fdSStefano Zampini } 3376d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3377d06fc5fdSStefano Zampini n_ISForEdges = 0; 3378d06fc5fdSStefano Zampini } 3379d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3380d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3381d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3382d06fc5fdSStefano Zampini } 3383d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3384d06fc5fdSStefano Zampini n_ISForFaces = 0; 3385d06fc5fdSStefano Zampini } 338670022509SStefano Zampini 338770022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 338870022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 338970022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 339070022509SStefano Zampini if (pcbddc->NullSpace) { 339170022509SStefano Zampini PetscBool tbool[2],gbool[2]; 339270022509SStefano Zampini 339370022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3394b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3395d06fc5fdSStefano Zampini if (!ISForEdges) { 3396d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3397d06fc5fdSStefano Zampini } 3398b8ffe317SStefano Zampini } 3399d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3400d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3401b2566f29SBarry Smith ierr = MPIU_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3402d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3403d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 340498a51de6SStefano Zampini } 340570022509SStefano Zampini #endif 340608122e43SStefano Zampini 3407674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3408674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3409674ae819SStefano Zampini if (nearnullsp) { 3410674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3411f4ddd8eeSStefano Zampini /* remove any stored info */ 3412f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3413f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3414f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3415f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3416f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3417473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3418f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3419f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3420f4ddd8eeSStefano Zampini } 3421984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3422984c4197SStefano Zampini nnsp_size = 0; 3423674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3424674ae819SStefano Zampini } 3425984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3426984c4197SStefano Zampini max_constraints = nnsp_size; 3427984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3428984c4197SStefano Zampini 3429674ae819SStefano Zampini /* 3430674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 34319162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 34329162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 34339162d606SStefano Zampini There can be multiple constraints per connected component 3434674ae819SStefano Zampini */ 3435674ae819SStefano Zampini n_vertices = 0; 3436674ae819SStefano Zampini if (ISForVertices) { 3437674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3438674ae819SStefano Zampini } 34399162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34409162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34419162d606SStefano Zampini 34429162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34439162d606SStefano Zampini total_counts *= max_constraints; 3444674ae819SStefano Zampini total_counts += n_vertices; 34454641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34469162d606SStefano Zampini 3447674ae819SStefano Zampini total_counts = 0; 3448674ae819SStefano Zampini max_size_of_constraint = 0; 3449674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34509162d606SStefano Zampini IS used_is; 3451674ae819SStefano Zampini if (i<n_ISForEdges) { 34529162d606SStefano Zampini used_is = ISForEdges[i]; 3453674ae819SStefano Zampini } else { 34549162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3455674ae819SStefano Zampini } 34569162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3457674ae819SStefano Zampini total_counts += j; 3458674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3459674ae819SStefano Zampini } 34609162d606SStefano 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); 34619162d606SStefano Zampini 3462984c4197SStefano Zampini /* get local part of global near null space vectors */ 3463785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3464984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3465984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3466e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3467e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3468984c4197SStefano Zampini } 3469674ae819SStefano Zampini 3470242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3471242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3472a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3473242a89d7SStefano Zampini 3474984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3475a773dcb8SStefano Zampini if (!skip_lapack) { 3476674ae819SStefano Zampini PetscScalar temp_work; 3477911cabfeSStefano Zampini 3478674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3479984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3480785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3481785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3482785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3483674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3484785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3485674ae819SStefano Zampini #endif 3486674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3487c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3488c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3489674ae819SStefano Zampini lwork = -1; 3490674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3491674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3492c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3493674ae819SStefano Zampini #else 3494c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3495674ae819SStefano Zampini #endif 3496674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3497984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3498674ae819SStefano Zampini #else /* on missing GESVD */ 3499674ae819SStefano Zampini /* SVD */ 3500674ae819SStefano Zampini PetscInt max_n,min_n; 3501674ae819SStefano Zampini max_n = max_size_of_constraint; 3502984c4197SStefano Zampini min_n = max_constraints; 3503984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3504674ae819SStefano Zampini min_n = max_size_of_constraint; 3505984c4197SStefano Zampini max_n = max_constraints; 3506674ae819SStefano Zampini } 3507785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3508674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3509785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3510674ae819SStefano Zampini #endif 3511674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3512674ae819SStefano Zampini lwork = -1; 3513e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3514e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3515b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3516674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3517674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35189162d606SStefano 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)); 3519674ae819SStefano Zampini #else 35209162d606SStefano 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)); 3521674ae819SStefano Zampini #endif 3522674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3523984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3524984c4197SStefano Zampini #endif /* on missing GESVD */ 3525674ae819SStefano Zampini /* Allocate optimal workspace */ 3526674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3527854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3528674ae819SStefano Zampini } 3529674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3530674ae819SStefano Zampini total_counts = 0; 35319162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 35329162d606SStefano Zampini constraints_data_ptr[0] = 0; 3533674ae819SStefano Zampini /* vertices */ 35349162d606SStefano Zampini if (n_vertices) { 3535674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 35369162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3537674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35389162d606SStefano Zampini constraints_n[total_counts] = 1; 35399162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35409162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35419162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3542674ae819SStefano Zampini total_counts++; 3543674ae819SStefano Zampini } 3544674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3545674ae819SStefano Zampini n_vertices = total_counts; 3546674ae819SStefano Zampini } 3547984c4197SStefano Zampini 3548674ae819SStefano Zampini /* edges and faces */ 35499162d606SStefano Zampini total_counts_cc = total_counts; 3550911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35519162d606SStefano Zampini IS used_is; 35529162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35539162d606SStefano Zampini 3554911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35559162d606SStefano Zampini used_is = ISForEdges[ncc]; 3556984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3557674ae819SStefano Zampini } else { 35589162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3559984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3560674ae819SStefano Zampini } 3561674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35629162d606SStefano Zampini 35639162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35649162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3565984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3566984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3567674ae819SStefano Zampini if (nnsp_has_cnst) { 35685b08dc53SStefano Zampini PetscScalar quad_value; 35699162d606SStefano Zampini 35709162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35719162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35729162d606SStefano Zampini 3573a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3574674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3575a773dcb8SStefano Zampini } else { 3576a773dcb8SStefano Zampini quad_value = 1.0; 3577a773dcb8SStefano Zampini } 3578674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35799162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3580674ae819SStefano Zampini } 35819162d606SStefano Zampini temp_constraints++; 3582674ae819SStefano Zampini total_counts++; 3583674ae819SStefano Zampini } 3584674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3585984c4197SStefano Zampini PetscReal real_value; 35869162d606SStefano Zampini PetscScalar *ptr_to_data; 35879162d606SStefano Zampini 3588984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35899162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3590674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35919162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3592674ae819SStefano Zampini } 3593984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3594984c4197SStefano Zampini /* check if array is null on the connected component */ 3595e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35969162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35975b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3598674ae819SStefano Zampini temp_constraints++; 3599674ae819SStefano Zampini total_counts++; 36009162d606SStefano Zampini if (!idxs_copied) { 36019162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 36029162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3603674ae819SStefano Zampini } 3604674ae819SStefano Zampini } 36059162d606SStefano Zampini } 36069162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 360745a1bb75SStefano Zampini valid_constraints = temp_constraints; 3608eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3609a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 36109162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 36119162d606SStefano Zampini 36129162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3613a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 36149162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3615a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 36169162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3617a773dcb8SStefano Zampini } else { /* perform SVD */ 3618984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 36199162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3620674ae819SStefano Zampini 3621674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3622984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3623984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3624984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3625984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3626984c4197SStefano Zampini from that computed using LAPACKgesvd 3627984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3628984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3629984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3630674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3631e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3632984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3633674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3634674ae819SStefano Zampini for (k=0;k<j+1;k++) { 36359162d606SStefano 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)); 3636674ae819SStefano Zampini } 3637674ae819SStefano Zampini } 3638e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3639e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3640e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3641674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3642c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3643674ae819SStefano Zampini #else 3644c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3645674ae819SStefano Zampini #endif 3646674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3647984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3648984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3649674ae819SStefano Zampini j = 0; 3650984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3651674ae819SStefano Zampini total_counts = total_counts-j; 365245a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3653e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3654c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3655c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3656c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3657c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3658c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3659c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3660674ae819SStefano Zampini if (j<temp_constraints) { 3661984c4197SStefano Zampini PetscInt ii; 3662984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3663674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36649162d606SStefano 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)); 3665674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3666984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3667674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36689162d606SStefano 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]; 3669674ae819SStefano Zampini } 3670674ae819SStefano Zampini } 3671674ae819SStefano Zampini } 3672674ae819SStefano Zampini #else /* on missing GESVD */ 3673e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3674e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3675b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3676674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3677674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36789162d606SStefano 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)); 3679674ae819SStefano Zampini #else 36809162d606SStefano 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)); 3681674ae819SStefano Zampini #endif 3682984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3683674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3684984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3685e310c8b4SStefano Zampini k = temp_constraints; 3686e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3687674ae819SStefano Zampini j = 0; 3688e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 368945a1bb75SStefano Zampini valid_constraints = k-j; 3690911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3691984c4197SStefano Zampini #endif /* on missing GESVD */ 3692674ae819SStefano Zampini } 3693a773dcb8SStefano Zampini } 36949162d606SStefano Zampini /* update pointers information */ 36959162d606SStefano Zampini if (valid_constraints) { 36969162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36979162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36989162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36999162d606SStefano Zampini /* set change_of_basis flag */ 370045a1bb75SStefano Zampini if (boolforchange) { 3701b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 37029162d606SStefano Zampini } 3703b3d85658SStefano Zampini total_counts_cc++; 370445a1bb75SStefano Zampini } 370545a1bb75SStefano Zampini } 3706984c4197SStefano Zampini /* free workspace */ 37078f1c130eSStefano Zampini if (!skip_lapack) { 3708984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3709984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3710984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3711984c4197SStefano Zampini #endif 3712984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3713984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3714984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3715984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3716984c4197SStefano Zampini #endif 3717984c4197SStefano Zampini } 3718984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3719984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3720984c4197SStefano Zampini } 3721984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3722cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3723cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3724cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3725cf5a6209SStefano Zampini } 3726cf5a6209SStefano Zampini if (n_ISForFaces) { 3727cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3728cf5a6209SStefano Zampini } 3729cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3730cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3731cf5a6209SStefano Zampini } 3732cf5a6209SStefano Zampini if (n_ISForEdges) { 3733cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3734cf5a6209SStefano Zampini } 3735cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 373608122e43SStefano Zampini } else { 373708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3738984c4197SStefano Zampini 373908122e43SStefano Zampini total_counts = 0; 374008122e43SStefano Zampini n_vertices = 0; 3741d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3742d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 374308122e43SStefano Zampini } 374408122e43SStefano Zampini max_constraints = 0; 37459162d606SStefano Zampini total_counts_cc = 0; 374608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 374708122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37489162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 374908122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 375008122e43SStefano Zampini } 37519162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37529162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37539162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37549162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 375574d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37569162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37579162d606SStefano Zampini total_counts_cc = 0; 37589162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37599162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37609162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 376108122e43SStefano Zampini } 376208122e43SStefano Zampini } 37639162d606SStefano Zampini #if 0 37649162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37659162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37669162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37679162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37689162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37699162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37709162d606SStefano Zampini } 37719162d606SStefano Zampini printf("\n"); 37729162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37739162d606SStefano Zampini } 37741b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37758bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37761b968477SStefano Zampini } 37771b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37788bec7fa6SStefano 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]); 37791b968477SStefano Zampini } 378008122e43SStefano Zampini #endif 378108122e43SStefano Zampini 37828bec7fa6SStefano Zampini max_size_of_constraint = 0; 37839162d606SStefano 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]); 37849162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 378508122e43SStefano Zampini /* Change of basis */ 3786b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 378708122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 378808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 378908122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3790b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 379108122e43SStefano Zampini } 379208122e43SStefano Zampini } 379308122e43SStefano Zampini } 379408122e43SStefano Zampini } 3795984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37964f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 379708122e43SStefano Zampini 37989162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37999162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 38006c4ed002SBarry 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); 3801674ae819SStefano Zampini 3802674ae819SStefano Zampini /* Create constraint matrix */ 3803674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 380416f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3805984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3806984c4197SStefano Zampini 3807984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3808a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3809a717540cSStefano Zampini qr_needed = PETSC_FALSE; 381074d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3811984c4197SStefano Zampini total_primal_vertices=0; 3812b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 38139162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 38149162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 381572b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 38169162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3817b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 381864efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 38199162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38209162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3821a717540cSStefano Zampini } 3822b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 382374d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3824a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3825a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3826a717540cSStefano Zampini } 3827fa434743SStefano Zampini } else { 3828b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3829fa434743SStefano Zampini } 3830a717540cSStefano Zampini } 3831b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3832b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3833674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 383470022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3835b3d85658SStefano Zampini 38364f1b2e48SStefano 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); 38370e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38380e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3839984c4197SStefano Zampini 3840984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 384174d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3842785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3843984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 384474d5cdf7SStefano Zampini 3845984c4197SStefano Zampini j = total_primal_vertices; 384674d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3847b3d85658SStefano Zampini cum = total_primal_vertices; 38489162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38494641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3850b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3851b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3852b3d85658SStefano Zampini cum++; 38539162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 385474d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 385574d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 385674d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 385774d5cdf7SStefano Zampini } 38589162d606SStefano Zampini j += constraints_n[i]; 3859674ae819SStefano Zampini } 3860674ae819SStefano Zampini } 3861674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3862674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3863088faed8SStefano Zampini 3864674ae819SStefano Zampini /* set values in constraint matrix */ 3865984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38660e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3867674ae819SStefano Zampini } 3868984c4197SStefano Zampini total_counts = total_primal_vertices; 38699162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38704641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38719162d606SStefano Zampini PetscInt *cols; 38729162d606SStefano Zampini 38739162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38749162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38759162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38769162d606SStefano Zampini PetscInt row = total_counts+k; 38779162d606SStefano Zampini PetscScalar *vals; 38789162d606SStefano Zampini 38799162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38809162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38819162d606SStefano Zampini } 38829162d606SStefano Zampini total_counts += constraints_n[i]; 3883674ae819SStefano Zampini } 3884674ae819SStefano Zampini } 3885674ae819SStefano Zampini /* assembling */ 3886674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3887674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3888088faed8SStefano Zampini 3889984c4197SStefano Zampini /* 38906a9046bcSBarry Smith ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3891984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3892f159cad9SBarry Smith ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr); 3893984c4197SStefano Zampini */ 3894674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3895674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3896026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3897984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3898984c4197SStefano Zampini /* working stuff for GEQRF */ 389981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3900984c4197SStefano Zampini PetscBLASInt lqr_work; 3901984c4197SStefano Zampini /* working stuff for UNGQR */ 3902984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3903984c4197SStefano Zampini PetscBLASInt lgqr_work; 3904984c4197SStefano Zampini /* working stuff for TRTRS */ 3905984c4197SStefano Zampini PetscScalar *trs_rhs; 39063f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3907984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3908984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3909984c4197SStefano Zampini PetscScalar *start_vals; 3910984c4197SStefano Zampini /* working stuff for values insertion */ 39114641a718SStefano Zampini PetscBT is_primal; 391264efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3913906d46d4SStefano Zampini /* matrix sizes */ 3914906d46d4SStefano Zampini PetscInt global_size,local_size; 3915906d46d4SStefano Zampini /* temporary change of basis */ 3916906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3917cf5a6209SStefano Zampini /* extra space for debugging */ 3918cf5a6209SStefano Zampini PetscScalar *dbg_work; 3919984c4197SStefano Zampini 3920906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3921906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 392216f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3923bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3924906d46d4SStefano Zampini /* nonzeros for local mat */ 3925bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 39261dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3927bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 39281dd7afcfSStefano Zampini } else { 39291dd7afcfSStefano Zampini const PetscInt *ii; 39301dd7afcfSStefano Zampini PetscInt n; 39311dd7afcfSStefano Zampini PetscBool flg_row; 39321dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39331dd7afcfSStefano Zampini for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i]; 39341dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);CHKERRQ(ierr); 39351dd7afcfSStefano Zampini } 39369162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3937a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39389162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3939a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39409162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3941a717540cSStefano Zampini } else { 39429162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39439162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3944a717540cSStefano Zampini } 3945a717540cSStefano Zampini } 3946a717540cSStefano Zampini } 3947906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3948bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 39491dd7afcfSStefano Zampini /* Set interior change in the matrix */ 39501dd7afcfSStefano Zampini if (!pcbddc->benign_change || pcbddc->fake_change) { 3951bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3952906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3953a717540cSStefano Zampini } 39541dd7afcfSStefano Zampini } else { 39551dd7afcfSStefano Zampini const PetscInt *ii,*jj; 39561dd7afcfSStefano Zampini PetscScalar *aa; 39571dd7afcfSStefano Zampini PetscInt n; 39581dd7afcfSStefano Zampini PetscBool flg_row; 39591dd7afcfSStefano Zampini ierr = MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39601dd7afcfSStefano Zampini ierr = MatSeqAIJGetArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39611dd7afcfSStefano Zampini for (i=0;i<n;i++) { 39621dd7afcfSStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);CHKERRQ(ierr); 39631dd7afcfSStefano Zampini } 39641dd7afcfSStefano Zampini ierr = MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);CHKERRQ(ierr); 39651dd7afcfSStefano Zampini ierr = MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);CHKERRQ(ierr); 39661dd7afcfSStefano Zampini } 3967a717540cSStefano Zampini 3968a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3969a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3970a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3971a717540cSStefano Zampini } 3972a717540cSStefano Zampini 3973a717540cSStefano Zampini 3974a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3975a717540cSStefano Zampini /* 3976a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3977a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3978a717540cSStefano Zampini 3979a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3980a717540cSStefano Zampini 3981a6b551f4SStefano 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) 3982a6b551f4SStefano Zampini 3983a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3984a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3985a717540cSStefano Zampini | ... | 3986a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3987a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3988a717540cSStefano Zampini 3989a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3990a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3991a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3992a6b551f4SStefano Zampini 3993a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3994a717540cSStefano Zampini */ 3995a717540cSStefano Zampini if (qr_needed) { 3996984c4197SStefano Zampini /* space to store Q */ 3997854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3998984c4197SStefano Zampini /* first we issue queries for optimal work */ 39993f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 40003f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 40013f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4002984c4197SStefano Zampini lqr_work = -1; 40033f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 4004984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 4005984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 4006785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 4007984c4197SStefano Zampini lgqr_work = -1; 40083f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 40093f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 40103f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 40113f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40123f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 40133f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 4014984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 4015984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 4016785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 4017984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 4018785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 4019984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 4020785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 4021a717540cSStefano Zampini /* allocating workspace for check */ 4022a717540cSStefano Zampini if (pcbddc->dbg_flag) { 4023cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 4024a717540cSStefano Zampini } 4025a717540cSStefano Zampini } 4026984c4197SStefano Zampini /* array to store whether a node is primal or not */ 40274641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 4028473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 40290e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 40306c4ed002SBarry 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); 403139e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 403239e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 403339e2fb2aSStefano Zampini } 403439e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 4035984c4197SStefano Zampini 4036a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 40379162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40389162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40394641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4040984c4197SStefano Zampini /* get constraint info */ 40419162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4042984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4043984c4197SStefano Zampini 4044984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40459162d606SStefano 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); 4046674ae819SStefano Zampini } 4047984c4197SStefano Zampini 4048fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4049a717540cSStefano Zampini 4050a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4051a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40529162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4053a717540cSStefano Zampini } 4054984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40559162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4056984c4197SStefano Zampini 4057984c4197SStefano Zampini /* compute QR decomposition of constraints */ 40583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40593f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4061674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40623f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4063984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4064674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4065984c4197SStefano Zampini 4066984c4197SStefano Zampini /* explictly compute R^-T */ 4067984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4068984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40693f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40703f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40713f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4073984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40743f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4075984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4076984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4077984c4197SStefano Zampini 4078a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40793f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40803f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40813f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40823f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4083984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40843f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4085984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4086984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4087984c4197SStefano Zampini 4088984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4089984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4090984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40923f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40933f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40953f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4097984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40989162d606SStefano 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)); 4099984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 41009162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4101984c4197SStefano Zampini 4102984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 41039162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4104984c4197SStefano Zampini /* insert cols for primal dofs */ 4105984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4106984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 41079162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4108906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4109984c4197SStefano Zampini } 4110984c4197SStefano Zampini /* insert cols for dual dofs */ 4111984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 41129162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4113984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 41149162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4115906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4116984c4197SStefano Zampini j++; 4117674ae819SStefano Zampini } 4118674ae819SStefano Zampini } 4119984c4197SStefano Zampini 4120984c4197SStefano Zampini /* check change of basis */ 4121984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4122984c4197SStefano Zampini PetscInt ii,jj; 4123984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4124c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4125c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4126c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4127c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4128c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4129c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4130984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4131cf5a6209SStefano 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)); 4132984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4133984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4134984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4135cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4136cf5a6209SStefano 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; 4137674ae819SStefano Zampini } 4138674ae819SStefano Zampini } 4139984c4197SStefano Zampini if (!valid_qr) { 414022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4141984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4142984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4143cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4144cf5a6209SStefano 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])); 4145674ae819SStefano Zampini } 4146cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4147cf5a6209SStefano 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])); 4148984c4197SStefano Zampini } 4149984c4197SStefano Zampini } 4150984c4197SStefano Zampini } 4151674ae819SStefano Zampini } else { 415222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4153674ae819SStefano Zampini } 4154674ae819SStefano Zampini } 4155a717540cSStefano Zampini } else { /* simple transformation block */ 4156a717540cSStefano Zampini PetscInt row,col; 4157a6b551f4SStefano Zampini PetscScalar val,norm; 4158a6b551f4SStefano Zampini 4159a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41609162d606SStefano 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)); 4161a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 41629162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 41639162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4164bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 41659162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4166906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41679162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4168a717540cSStefano Zampini } else { 4169a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41709162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4171a717540cSStefano Zampini if (row != col) { 41729162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4173a717540cSStefano Zampini } else { 41749162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4175a717540cSStefano Zampini } 4176906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4177a717540cSStefano Zampini } 4178a717540cSStefano Zampini } 4179a717540cSStefano Zampini } 418098a51de6SStefano Zampini if (pcbddc->dbg_flag) { 418122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4182a717540cSStefano Zampini } 4183674ae819SStefano Zampini } 4184984c4197SStefano Zampini } else { 4185984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41869162d606SStefano 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); 4187674ae819SStefano Zampini } 4188674ae819SStefano Zampini } 4189674ae819SStefano Zampini } 4190a717540cSStefano Zampini 4191a717540cSStefano Zampini /* free workspace */ 4192a717540cSStefano Zampini if (qr_needed) { 4193984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4194cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4195984c4197SStefano Zampini } 4196984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4197984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4198984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4199984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4200984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4201674ae819SStefano Zampini } 4202a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4203906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4204906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4205906d46d4SStefano Zampini 4206906d46d4SStefano Zampini /* assembling of global change of variable */ 420788c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4208bbb9e6c6SStefano Zampini Mat tmat; 420916f15bc4SStefano Zampini PetscInt bs; 421016f15bc4SStefano Zampini 4211906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4212906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4213bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4214bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4215bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4216bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 421716f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 421816f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4219906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4220bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4221bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4222bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4223bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4224bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4225e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4226e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4227bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4228bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 422988c03ad3SStefano Zampini 4230906d46d4SStefano Zampini /* check */ 4231906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4232906d46d4SStefano Zampini PetscReal error; 4233906d46d4SStefano Zampini Vec x,x_change; 4234906d46d4SStefano Zampini 4235906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4236906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4237906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4238906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4239e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4240e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4241bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4242e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4243e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4244906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4245906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4246906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4247906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4248bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4249906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4250906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4251906d46d4SStefano Zampini } 4252b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4253b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4254b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4255bf3a8328SStefano Zampini 4256bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4257bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4258bf3a8328SStefano Zampini } 4259b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4260ac632422SStefano Zampini Mat S_new,tmat; 4261bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4262bbb9e6c6SStefano Zampini 4263bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 42646816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4265bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4266bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4267bf3a8328SStefano Zampini IS is_V; 4268b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4269b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4270b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4271b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4272b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4273bf3a8328SStefano Zampini } 4274bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4275ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4276b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4277ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4278bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4279bf3a8328SStefano Zampini const PetscScalar *array; 4280bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4281bf3a8328SStefano Zampini PetscInt i,n_V; 4282bf3a8328SStefano Zampini 4283b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4284b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4285b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4286b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4287b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4288b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4289b087196eSStefano Zampini PetscScalar val; 4290b087196eSStefano Zampini PetscInt idx; 4291b087196eSStefano Zampini 4292b087196eSStefano Zampini idx = idxs_V[i]; 4293b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4294b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4295b087196eSStefano Zampini } 4296b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4297b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4298bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4299bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4300bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4301bf3a8328SStefano Zampini } 4302ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4303ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4304ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4305ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4306b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4307ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4308bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4309b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4310bf3a8328SStefano Zampini } 4311ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4312ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4313ac632422SStefano Zampini } 4314b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 431588c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4316b96c3477SStefano Zampini } 4317c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4318c9db6a07SStefano Zampini if (sub_schurs->change) { 4319c9db6a07SStefano Zampini PetscInt i; 4320c9db6a07SStefano Zampini 4321c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4322c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4323c9db6a07SStefano Zampini } 4324c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4325c9db6a07SStefano Zampini } 4326b96c3477SStefano Zampini } 432716909a7fSStefano Zampini if (pcbddc->switch_static) { /* need to save the local change */ 432816909a7fSStefano Zampini pcbddc->switch_static_change = localChangeOfBasisMatrix; 432916909a7fSStefano Zampini } else { 4330906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 433116909a7fSStefano Zampini } 43321dd7afcfSStefano Zampini /* determine if any process has changed the pressures locally */ 43331dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43341dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43351dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->change_interior,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43361dd7afcfSStefano Zampini } 433772b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 433872b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 433972b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 434072b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 434172b8c272SStefano Zampini } 43421dd7afcfSStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) { 43431dd7afcfSStefano Zampini PetscBool needglobal = PETSC_FALSE; 43441dd7afcfSStefano Zampini if (pcbddc->benign_saddle_point) { 43451dd7afcfSStefano Zampini PetscBool have_null = (PetscBool)!!pcbddc->benign_change; 43461dd7afcfSStefano Zampini ierr = MPI_Allreduce(&have_null,&needglobal,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 43471dd7afcfSStefano Zampini } 43481dd7afcfSStefano Zampini if (!needglobal && pcbddc->user_ChangeOfBasisMatrix) { 4349b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4350b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4351906d46d4SStefano Zampini } else { 43521dd7afcfSStefano Zampini Mat benign_global = NULL; 43531dd7afcfSStefano Zampini if (needglobal) { 43541dd7afcfSStefano Zampini Mat tmat; 43551dd7afcfSStefano Zampini 43561dd7afcfSStefano Zampini pcbddc->change_interior = PETSC_TRUE; 43571dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 43581dd7afcfSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 43591dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43601dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 43611dd7afcfSStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 43621dd7afcfSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43631dd7afcfSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 43641dd7afcfSStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 43651dd7afcfSStefano Zampini if (pcbddc->benign_change) { 43661dd7afcfSStefano Zampini Mat M; 43671dd7afcfSStefano Zampini 43681dd7afcfSStefano Zampini ierr = MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);CHKERRQ(ierr); 43691dd7afcfSStefano Zampini ierr = MatDiagonalScale(M,pcis->vec1_N,NULL);CHKERRQ(ierr); 43701dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,M);CHKERRQ(ierr); 43711dd7afcfSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 4372906d46d4SStefano Zampini } else { 43731dd7afcfSStefano Zampini Mat eye; 43741dd7afcfSStefano Zampini PetscScalar *array; 43751dd7afcfSStefano Zampini 43761dd7afcfSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43771dd7afcfSStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&eye);CHKERRQ(ierr); 43781dd7afcfSStefano Zampini for (i=0;i<pcis->n;i++) { 43791dd7afcfSStefano Zampini ierr = MatSetValue(eye,i,i,array[i],INSERT_VALUES);CHKERRQ(ierr); 4380906d46d4SStefano Zampini } 43811dd7afcfSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 43821dd7afcfSStefano Zampini ierr = MatAssemblyBegin(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43831dd7afcfSStefano Zampini ierr = MatAssemblyEnd(eye,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 43841dd7afcfSStefano Zampini ierr = MatISSetLocalMat(tmat,eye);CHKERRQ(ierr); 43851dd7afcfSStefano Zampini ierr = MatDestroy(&eye);CHKERRQ(ierr); 43861dd7afcfSStefano Zampini } 43871dd7afcfSStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_INITIAL_MATRIX,&benign_global);CHKERRQ(ierr); 43881dd7afcfSStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 43891dd7afcfSStefano Zampini } 43901dd7afcfSStefano Zampini if (pcbddc->user_ChangeOfBasisMatrix) { 43911dd7afcfSStefano Zampini ierr = MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 43921dd7afcfSStefano Zampini ierr = MatDestroy(&benign_global);CHKERRQ(ierr); 43931dd7afcfSStefano Zampini } else if (needglobal) { 43941dd7afcfSStefano Zampini pcbddc->ChangeOfBasisMatrix = benign_global; 43951dd7afcfSStefano Zampini } 43961dd7afcfSStefano Zampini } 439716909a7fSStefano Zampini if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */ 439816909a7fSStefano Zampini IS is_global; 439916909a7fSStefano Zampini const PetscInt *gidxs; 440016909a7fSStefano Zampini 440116909a7fSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 440216909a7fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);CHKERRQ(ierr); 440316909a7fSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);CHKERRQ(ierr); 440416909a7fSStefano Zampini ierr = MatGetSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);CHKERRQ(ierr); 440516909a7fSStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 440616909a7fSStefano Zampini } 44071dd7afcfSStefano Zampini } 44081dd7afcfSStefano Zampini if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) { 44091dd7afcfSStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&pcbddc->work_change);CHKERRQ(ierr); 4410b9b85e73SStefano Zampini } 4411a717540cSStefano Zampini 441272b8c272SStefano Zampini if (!pcbddc->fake_change) { 44134f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 44144f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 44154f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 44164f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4417019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4418019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4419019a44ceSStefano Zampini pcbddc->local_primal_size++; 4420019a44ceSStefano Zampini } 4421019a44ceSStefano Zampini 4422019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4423727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4424727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 44259f47a83aSStefano 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); 4426c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 44270e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 44289f47a83aSStefano 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); 4429727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4430727cdba6SStefano Zampini } 44310e6343abSStefano Zampini } 4432727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4433b2566f29SBarry Smith ierr = MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 443472b8c272SStefano Zampini } 443572b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4436727cdba6SStefano Zampini 4437a717540cSStefano Zampini /* flush dbg viewer */ 4438b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4439b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4440b8ffe317SStefano Zampini } 4441a717540cSStefano Zampini 4442e310c8b4SStefano Zampini /* free workspace */ 4443a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 44444641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 444508122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 44469162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 44479162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 444808122e43SStefano Zampini } else { 44499162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 44509162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 44519162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 445208122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 445308122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 44549162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 44559162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 445608122e43SStefano Zampini } 4457674ae819SStefano Zampini PetscFunctionReturn(0); 4458674ae819SStefano Zampini } 4459674ae819SStefano Zampini 4460674ae819SStefano Zampini #undef __FUNCT__ 4461674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4462674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4463674ae819SStefano Zampini { 4464674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4465674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4466674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 446714f95afaSStefano Zampini PetscInt ierr,i,N; 4468674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4469674ae819SStefano Zampini 4470674ae819SStefano Zampini PetscFunctionBegin; 44718e61c736SStefano Zampini /* Reset previously computed graph */ 44728e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4473674ae819SStefano Zampini /* Init local Graph struct */ 44747fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 44753bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4476674ae819SStefano Zampini 4477575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 44785099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 44795099eff2SStefano 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); 4480575ad6abSStefano Zampini } 44819577ea80SStefano Zampini 4482674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4483d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 44844d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 44854d379d7bSStefano Zampini PetscInt nvtxs; 4486e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4487674ae819SStefano Zampini 44882fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 44892fffb893SStefano Zampini if (flg_row) { 44904d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4491b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 44922fffb893SStefano Zampini } 44932fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4494674ae819SStefano Zampini } 44959b28b941SStefano Zampini if (pcbddc->dbg_flag) { 44969b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4497674ae819SStefano Zampini } 4498674ae819SStefano Zampini 449963602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 450063602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 450163602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 450295ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 450363602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4504e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 450563602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4506674ae819SStefano Zampini } 450763602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 450863602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 450963602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4510674ae819SStefano Zampini } 451163602bcaSStefano Zampini } else { 451263602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 451363602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4514854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 451563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 451663602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 451763602bcaSStefano Zampini } 451863602bcaSStefano Zampini } 4519674ae819SStefano Zampini } 4520674ae819SStefano Zampini 4521674ae819SStefano Zampini /* Setup of Graph */ 4522785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4523e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4524785d1243SStefano Zampini } 4525785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4526e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4527785d1243SStefano Zampini } 452830368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 452930368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 453030368db7SStefano Zampini } 453114f95afaSStefano 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); 4532674ae819SStefano Zampini 45334f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 45344f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 45354f1b2e48SStefano Zampini PetscInt *local_subs; 45364f1b2e48SStefano Zampini 45374f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 45384f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 45394f1b2e48SStefano Zampini const PetscInt *idxs; 45404f1b2e48SStefano Zampini PetscInt nl,j; 45414f1b2e48SStefano Zampini 45424f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 45434f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45444f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 45454f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 45464f1b2e48SStefano Zampini } 45474f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 45484f1b2e48SStefano Zampini } 45494f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 45504f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 45514f1b2e48SStefano Zampini } 45524f1b2e48SStefano Zampini 4553674ae819SStefano Zampini /* Graph's connected components analysis */ 4554674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4555674ae819SStefano Zampini 4556674ae819SStefano Zampini /* print some info to stdout */ 4557674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4558302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4559674ae819SStefano Zampini } 4560674ae819SStefano Zampini PetscFunctionReturn(0); 4561674ae819SStefano Zampini } 4562674ae819SStefano Zampini 4563dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4564674ae819SStefano Zampini #undef __FUNCT__ 4565674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4566dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4567674ae819SStefano Zampini { 4568dc456d91SStefano Zampini PetscSF sf; 4569dc456d91SStefano Zampini PetscLayout map; 4570dc456d91SStefano Zampini const PetscInt *idxs; 4571dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4572dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4573dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4574dc456d91SStefano Zampini PetscMPIInt commsize; 4575674ae819SStefano Zampini PetscBool first_found; 4576674ae819SStefano Zampini PetscErrorCode ierr; 4577674ae819SStefano Zampini 4578674ae819SStefano Zampini PetscFunctionBegin; 4579dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4580dc456d91SStefano Zampini if (subset_mult) { 4581dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4582dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4583dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4584674ae819SStefano Zampini } 4585dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4586dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4587dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4588dc456d91SStefano Zampini for (i=0;i<n;i++) { 4589dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4590dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4591674ae819SStefano Zampini } 4592dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4593b2566f29SBarry Smith ierr = MPIU_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4594dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4595dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4596dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4597dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4598dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4599dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4600dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4601dc456d91SStefano Zampini 4602dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4603dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4604dc456d91SStefano Zampini if (subset_mult) { 4605dc456d91SStefano Zampini const PetscInt* idxs_mult; 4606dc456d91SStefano Zampini 4607dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4608dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4609dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4610674ae819SStefano Zampini } else { 4611dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4612674ae819SStefano Zampini } 4613dc456d91SStefano Zampini /* local size of new subset */ 4614dc456d91SStefano Zampini n_n = 0; 4615dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4616dc456d91SStefano Zampini 4617dc456d91SStefano Zampini /* global indexes in layout */ 4618dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4619dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4620dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4621dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4622dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4623dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4624dc456d91SStefano Zampini 4625dc456d91SStefano Zampini /* reduce from leaves to roots */ 4626dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 462764a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 462864a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4629dc456d91SStefano Zampini 4630dc456d91SStefano Zampini /* count indexes in local part of layout */ 4631674ae819SStefano Zampini nlocals = 0; 4632674ae819SStefano Zampini first_index = -1; 4633674ae819SStefano Zampini first_found = PETSC_FALSE; 4634dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4635dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4636674ae819SStefano Zampini first_found = PETSC_TRUE; 4637674ae819SStefano Zampini first_index = i; 4638674ae819SStefano Zampini } 4639dc456d91SStefano Zampini nlocals += root_data[i]; 4640674ae819SStefano Zampini } 4641dc456d91SStefano Zampini 4642dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 46435fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4644dc456d91SStefano Zampini start = 0; 464564a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46465fa240b1SStefano Zampini #else 464764a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 46485fa240b1SStefano Zampini start = start-nlocals; 46495fa240b1SStefano Zampini #endif 46505fa240b1SStefano Zampini 4651dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4652dc456d91SStefano Zampini *N_n = start + nlocals; 4653dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4654dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4655674ae819SStefano Zampini } 46565fa240b1SStefano Zampini 46575fa240b1SStefano Zampini /* adapt root data with cumulative */ 4658674ae819SStefano Zampini if (first_found) { 4659dc456d91SStefano Zampini PetscInt old_index; 4660dc456d91SStefano Zampini 4661dc456d91SStefano Zampini root_data[first_index] += start; 4662674ae819SStefano Zampini old_index = first_index; 4663dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4664dc456d91SStefano Zampini if (root_data[i]) { 4665dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4666674ae819SStefano Zampini old_index = i; 4667674ae819SStefano Zampini } 4668674ae819SStefano Zampini } 4669674ae819SStefano Zampini } 4670dc456d91SStefano Zampini 4671dc456d91SStefano Zampini /* from roots to leaves */ 4672dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4673dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4674dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4675dc456d91SStefano Zampini 4676dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4677dc456d91SStefano Zampini if (subset_mult) { 4678dc456d91SStefano Zampini const PetscInt* idxs_mult; 4679dc456d91SStefano Zampini PetscInt cum; 4680dc456d91SStefano Zampini 4681dc456d91SStefano Zampini cum = 0; 4682dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4683dc456d91SStefano Zampini for (i=0;i<n;i++) { 4684dc456d91SStefano Zampini PetscInt j; 4685dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4686674ae819SStefano Zampini } 4687dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4688674ae819SStefano Zampini } else { 4689dc456d91SStefano Zampini for (i=0;i<n;i++) { 4690dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4691674ae819SStefano Zampini } 4692674ae819SStefano Zampini } 4693dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4694dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4695674ae819SStefano Zampini PetscFunctionReturn(0); 4696674ae819SStefano Zampini } 46979a7d3425SStefano Zampini 46989a7d3425SStefano Zampini #undef __FUNCT__ 46999a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 47009a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 47019a7d3425SStefano Zampini { 47029a7d3425SStefano Zampini PetscInt i,j; 47039a7d3425SStefano Zampini PetscScalar *alphas; 47049a7d3425SStefano Zampini PetscErrorCode ierr; 47059a7d3425SStefano Zampini 47069a7d3425SStefano Zampini PetscFunctionBegin; 47079a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4708785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 47099a7d3425SStefano Zampini for (i=0;i<n;i++) { 47109a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 47119a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 47129a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 47139a7d3425SStefano Zampini } 47149a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 47159a7d3425SStefano Zampini PetscFunctionReturn(0); 47169a7d3425SStefano Zampini } 47179a7d3425SStefano Zampini 4718e7931f94SStefano Zampini #undef __FUNCT__ 471970cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4720b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4721e7931f94SStefano Zampini { 472252e5ac9dSStefano Zampini IS ranks_send_to; 4723e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4724e7931f94SStefano Zampini PetscMPIInt size,rank,color; 472552e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 472652e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 47273837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 47282b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4729e7931f94SStefano Zampini PetscSubcomm subcomm; 473052e5ac9dSStefano Zampini PetscErrorCode ierr; 4731a57a6d2fSStefano Zampini 4732e7931f94SStefano Zampini PetscFunctionBegin; 4733c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 4734c5929fdfSBarry Smith ierr = PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 4735c5929fdfSBarry Smith ierr = PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4736e7931f94SStefano Zampini 4737e7931f94SStefano Zampini /* Get info on mapping */ 47383bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 47393bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4740e7931f94SStefano Zampini 4741e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4742785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4743e7931f94SStefano Zampini xadj[0] = 0; 4744e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4745785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4746785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4747e7931f94SStefano Zampini 47482b510759SStefano Zampini if (threshold) { 4749d023bfaeSStefano Zampini PetscInt xadj_count = 0; 47502b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4751d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4752d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4753d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4754d023bfaeSStefano Zampini xadj_count++; 4755e7931f94SStefano Zampini } 4756e7931f94SStefano Zampini } 4757d023bfaeSStefano Zampini xadj[1] = xadj_count; 4758c8587f34SStefano Zampini } else { 4759e7931f94SStefano Zampini if (xadj[1]) { 4760e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4761e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4762c8587f34SStefano Zampini } 4763e7931f94SStefano Zampini } 47643bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4765e7931f94SStefano Zampini if (use_square) { 4766e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4767e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4768e7931f94SStefano Zampini } 4769e7931f94SStefano Zampini } 4770e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4771e7931f94SStefano Zampini 47723837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4773e7931f94SStefano Zampini 4774e7931f94SStefano Zampini /* 4775e7931f94SStefano Zampini Restrict work on active processes only. 4776e7931f94SStefano Zampini */ 4777e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4778e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4779e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 47802b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4781d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4782e7931f94SStefano Zampini if (color) { 4783e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4784e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4785e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4786c8587f34SStefano Zampini } else { 478752e5ac9dSStefano Zampini Mat subdomain_adj; 478852e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 478952e5ac9dSStefano Zampini MatPartitioning partitioner; 479052e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 479152e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4792b0c7d250SStefano Zampini PetscBool aggregate; 4793b0c7d250SStefano Zampini 4794306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4795785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4796e7931f94SStefano Zampini prank = rank; 4797306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 47988002ef2cSStefano Zampini /* 4799e7931f94SStefano Zampini for (i=0;i<size;i++) { 4800e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4801c8587f34SStefano Zampini } 48028002ef2cSStefano Zampini */ 4803e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4804e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4805c8587f34SStefano Zampini } 4806e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4807b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4808b0c7d250SStefano Zampini if (aggregate) { 4809b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4810b0c7d250SStefano Zampini PetscMPIInt nrank; 4811b0c7d250SStefano Zampini PetscScalar *vals; 4812b0c7d250SStefano Zampini 4813b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4814b0c7d250SStefano Zampini lrows = 0; 4815b0c7d250SStefano Zampini if (nrank<redprocs) { 4816b0c7d250SStefano Zampini lrows = size/redprocs; 4817b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4818b0c7d250SStefano Zampini } 48195fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4820b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4821b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4822b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4823b0c7d250SStefano Zampini row = nrank; 4824b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4825b0c7d250SStefano Zampini cols = adjncy; 4826b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4827b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4828b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4829b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4830b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 483152e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 483252e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 483352e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4834b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4835b0c7d250SStefano Zampini } else { 4836306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4837b0c7d250SStefano Zampini } 483822b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4839e7931f94SStefano Zampini 4840e7931f94SStefano Zampini /* Partition */ 4841306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4842e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4843e7931f94SStefano Zampini if (use_vwgt) { 48443837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4845e7931f94SStefano Zampini v_wgt[0] = local_size; 4846e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4847c8587f34SStefano Zampini } 484828143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 484928143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4850e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4851e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 485222b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4853e7931f94SStefano Zampini 485452e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 485552e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 485652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 485752e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4858b0c7d250SStefano Zampini if (!redprocs) { 4859b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 486028143c3dSStefano Zampini } else { 4861b0c7d250SStefano Zampini PetscInt idxs[1]; 4862b0c7d250SStefano Zampini PetscMPIInt tag; 4863b0c7d250SStefano Zampini MPI_Request *reqs; 4864b0c7d250SStefano Zampini 4865b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4866b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4867b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4868b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 486928143c3dSStefano Zampini } 4870b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4871b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4872b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4873b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4874e7931f94SStefano Zampini } 487552e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4876e7931f94SStefano Zampini /* clean up */ 4877e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 487852e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4879e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4880e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4881e7931f94SStefano Zampini } 4882e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4883e7931f94SStefano Zampini 4884e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4885e7931f94SStefano Zampini i = 1; 4886e7931f94SStefano Zampini if (color) i=0; 4887e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4888e7931f94SStefano Zampini /* get back IS */ 4889e7931f94SStefano Zampini *is_sends = ranks_send_to; 4890e7931f94SStefano Zampini PetscFunctionReturn(0); 4891e7931f94SStefano Zampini } 4892e7931f94SStefano Zampini 4893e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4894e7931f94SStefano Zampini 4895e7931f94SStefano Zampini #undef __FUNCT__ 4896e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 489753a05cb3SStefano 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[]) 4898e7931f94SStefano Zampini { 489970cf5478SStefano Zampini Mat local_mat; 4900e7931f94SStefano Zampini IS is_sends_internal; 49019d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 490228143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 49039d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4904e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4905e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4906e7931f94SStefano Zampini const PetscInt* is_indices; 4907e7931f94SStefano Zampini MatType new_local_type; 4908e7931f94SStefano Zampini /* buffers */ 4909e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 491028143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 49119d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4912e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4913e7931f94SStefano Zampini /* MPI */ 491428143c3dSStefano Zampini MPI_Comm comm,comm_n; 491528143c3dSStefano Zampini PetscSubcomm subcomm; 4916e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 491728143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 491828143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 491928143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 492028143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 492128143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4922e7931f94SStefano Zampini PetscErrorCode ierr; 4923e7931f94SStefano Zampini 4924e7931f94SStefano Zampini PetscFunctionBegin; 492528143c3dSStefano Zampini /* TODO: add missing checks */ 492628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 492728143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 492828143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 492928143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4930e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 493128143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4932e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4933e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4934e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4935e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4936e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 493728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 493870cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 493970cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 494028143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 494170cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 494270cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 494370cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 494470cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 494570cf5478SStefano Zampini } 4946e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4947e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4948e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4949e7931f94SStefano Zampini if (!is_sends) { 495028143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4951b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4952c8587f34SStefano Zampini } else { 4953e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4954e7931f94SStefano Zampini is_sends_internal = is_sends; 4955c8587f34SStefano Zampini } 4956e7931f94SStefano Zampini 4957e7931f94SStefano Zampini /* get comm */ 4958a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4959e7931f94SStefano Zampini 4960e7931f94SStefano Zampini /* compute number of sends */ 4961e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4962e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4963e7931f94SStefano Zampini 4964e7931f94SStefano Zampini /* compute number of receives */ 4965e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4966785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4967e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4968e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4969e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4970e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4971e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4972e7931f94SStefano Zampini 497328143c3dSStefano Zampini /* restrict comm if requested */ 497428143c3dSStefano Zampini subcomm = 0; 497528143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 497628143c3dSStefano Zampini if (restrict_comm) { 4977779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4978779c1cceSStefano Zampini 497928143c3dSStefano Zampini color = 0; 498053a05cb3SStefano Zampini if (restrict_full) { 498153a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 498253a05cb3SStefano Zampini } else { 498353a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 498453a05cb3SStefano Zampini } 4985b2566f29SBarry Smith ierr = MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 498628143c3dSStefano Zampini subcommsize = commsize - subcommsize; 498728143c3dSStefano Zampini /* check if reuse has been requested */ 498828143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 498928143c3dSStefano Zampini if (*mat_n) { 499028143c3dSStefano Zampini PetscMPIInt subcommsize2; 499128143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 499228143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 499328143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 499428143c3dSStefano Zampini } else { 499528143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 499628143c3dSStefano Zampini } 499728143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4998779c1cceSStefano Zampini PetscMPIInt rank; 4999779c1cceSStefano Zampini 5000779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 500128143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 500228143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 500328143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5004306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 500528143c3dSStefano Zampini } 500628143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 500728143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 500828143c3dSStefano Zampini } else { 500928143c3dSStefano Zampini comm_n = comm; 501028143c3dSStefano Zampini } 501128143c3dSStefano Zampini 5012e7931f94SStefano Zampini /* prepare send/receive buffers */ 5013785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 5014e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 5015785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 5016e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 501728143c3dSStefano Zampini if (nis) { 5018854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 501928143c3dSStefano Zampini } 5020e7931f94SStefano Zampini 502128143c3dSStefano Zampini /* Get data from local matrices */ 50226c4ed002SBarry Smith if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 5023e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 5024e7931f94SStefano Zampini /* 5025e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 5026e7931f94SStefano Zampini send_buffer_idxs should contain: 5027e7931f94SStefano Zampini - MatType_PRIVATE type 5028e7931f94SStefano Zampini - PetscInt size_of_l2gmap 5029e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 5030e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 5031e7931f94SStefano Zampini */ 50326c4ed002SBarry Smith else { 5033e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 50343bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 5035854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 5036e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 5037e7931f94SStefano Zampini send_buffer_idxs[1] = i; 50383bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5039e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 50403bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 5041e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 5042e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5043e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 5044e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 5045c8587f34SStefano Zampini } 5046c8587f34SStefano Zampini } 5047e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 504828143c3dSStefano Zampini /* additional is (if any) */ 504928143c3dSStefano Zampini if (nis) { 505028143c3dSStefano Zampini PetscMPIInt psum; 505128143c3dSStefano Zampini PetscInt j; 505228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 505328143c3dSStefano Zampini PetscInt plen; 505428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 505528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 505628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 505728143c3dSStefano Zampini } 5058854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 505928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 506028143c3dSStefano Zampini PetscInt plen; 506128143c3dSStefano Zampini const PetscInt *is_array_idxs; 506228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 506328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 506428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 506528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 506628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 506728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 506828143c3dSStefano Zampini } 506928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 507028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 507128143c3dSStefano Zampini } 507228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 507328143c3dSStefano Zampini } 507428143c3dSStefano Zampini 5075e7931f94SStefano Zampini buf_size_idxs = 0; 5076e7931f94SStefano Zampini buf_size_vals = 0; 507728143c3dSStefano Zampini buf_size_idxs_is = 0; 5078e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5079e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5080e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 508128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5082e7931f94SStefano Zampini } 5083785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5084785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 508595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5086e7931f94SStefano Zampini 5087e7931f94SStefano Zampini /* get new tags for clean communications */ 5088e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5089e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 509028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5091e7931f94SStefano Zampini 5092e7931f94SStefano Zampini /* allocate for requests */ 5093785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5094785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 509595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5096785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5097785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 509895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5099e7931f94SStefano Zampini 5100e7931f94SStefano Zampini /* communications */ 5101e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5102e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 510328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5104e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5105e7931f94SStefano Zampini source_dest = onodes[i]; 5106e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5107e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5108e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5109e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 511028143c3dSStefano Zampini if (nis) { 511128143c3dSStefano 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); 511228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 511328143c3dSStefano Zampini } 5114e7931f94SStefano Zampini } 5115e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5116e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5117e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5118e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 511928143c3dSStefano Zampini if (nis) { 512028143c3dSStefano 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); 512128143c3dSStefano Zampini } 5122e7931f94SStefano Zampini } 5123e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5124e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5125e7931f94SStefano Zampini 5126e7931f94SStefano Zampini /* assemble new l2g map */ 5127e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5128e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51299d30be91SStefano Zampini new_local_rows = 0; 5130e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51319d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5132e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5133e7931f94SStefano Zampini } 51349d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5135e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 51369d30be91SStefano Zampini new_local_rows = 0; 5137e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 51389d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 51399d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5140e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5141e7931f94SStefano Zampini } 51429d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 51439d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5144e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5145e7931f94SStefano Zampini 5146e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5147e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5148e7931f94SStefano 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) */ 5149e7931f94SStefano Zampini if (n_recvs) { 515028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5151e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5152e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5153e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5154e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5155e7931f94SStefano Zampini break; 5156e7931f94SStefano Zampini } 5157e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5158e7931f94SStefano Zampini } 5159e7931f94SStefano Zampini switch (new_local_type_private) { 516028143c3dSStefano Zampini case MATDENSE_PRIVATE: 516128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5162e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5163e7931f94SStefano Zampini bs = 1; 516428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 516528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 516628143c3dSStefano Zampini bs = 1; 516728143c3dSStefano Zampini } 5168e7931f94SStefano Zampini break; 5169e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5170e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5171e7931f94SStefano Zampini bs = 1; 5172e7931f94SStefano Zampini break; 5173e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5174e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5175e7931f94SStefano Zampini break; 5176e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5177e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5178e7931f94SStefano Zampini break; 5179e7931f94SStefano Zampini default: 51809d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5181e7931f94SStefano Zampini break; 5182e7931f94SStefano Zampini } 518328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 518428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 518528143c3dSStefano Zampini bs = 1; 5186e7931f94SStefano Zampini } 5187e7931f94SStefano Zampini 518870cf5478SStefano Zampini /* create MATIS object if needed */ 518970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5190e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5191e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 519270cf5478SStefano Zampini } else { 519370cf5478SStefano Zampini /* it also destroys the local matrices */ 519470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 519570cf5478SStefano Zampini } 519670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5197e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 51989d30be91SStefano Zampini 51999d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 52009d30be91SStefano Zampini 52019d30be91SStefano Zampini /* Global to local map of received indices */ 52029d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 52039d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 52049d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 52059d30be91SStefano Zampini 52069d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 52079d30be91SStefano Zampini buf_size_idxs = 0; 52089d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 52099d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 52109d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 52119d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 52129d30be91SStefano Zampini } 52139d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 52149d30be91SStefano Zampini 52159d30be91SStefano Zampini /* set preallocation */ 52169d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 52179d30be91SStefano Zampini if (!newisdense) { 52189d30be91SStefano Zampini PetscInt *new_local_nnz=0; 52199d30be91SStefano Zampini 52209d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 52219d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 52229d30be91SStefano Zampini if (n_recvs) { 52239d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 52249d30be91SStefano Zampini } 52259d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 52269d30be91SStefano Zampini PetscInt j; 52279d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 52289d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 52299d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 52309d30be91SStefano Zampini } 52319d30be91SStefano Zampini } else { 52329d30be91SStefano Zampini /* TODO */ 52339d30be91SStefano Zampini } 52349d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 52359d30be91SStefano Zampini } 52369d30be91SStefano Zampini if (new_local_nnz) { 52379d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 52389d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 52399d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 52409d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52419d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 52429d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 52439d30be91SStefano Zampini } else { 52449d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52459d30be91SStefano Zampini } 52469d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 52479d30be91SStefano Zampini } else { 52489d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 52499d30be91SStefano Zampini } 5250e7931f94SStefano Zampini 5251e7931f94SStefano Zampini /* set values */ 5252e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 52539d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5254e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5255e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5256e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 52579d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5258e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5259e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5260e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 526128143c3dSStefano Zampini } else { 526228143c3dSStefano Zampini /* TODO */ 5263e7931f94SStefano Zampini } 5264e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5265e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5266e7931f94SStefano Zampini } 5267e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5268e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 526970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 527070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 52719d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 52729d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5273e7931f94SStefano Zampini 5274dfd14d43SStefano Zampini #if 0 527528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5276e7931f94SStefano Zampini Vec lvec,rvec; 5277e7931f94SStefano Zampini PetscReal infty_error; 5278e7931f94SStefano Zampini 52792a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5280e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5281e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5282e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 528370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5284e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5285e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5286e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5287e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5288e7931f94SStefano Zampini } 528928143c3dSStefano Zampini #endif 5290e7931f94SStefano Zampini 529128143c3dSStefano Zampini /* assemble new additional is (if any) */ 529228143c3dSStefano Zampini if (nis) { 529328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 529428143c3dSStefano Zampini 529528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5296854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 529728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 529828143c3dSStefano Zampini psum = 0; 529928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 530028143c3dSStefano Zampini for (j=0;j<nis;j++) { 530128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 530228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 530328143c3dSStefano Zampini psum += plen; 530428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 530528143c3dSStefano Zampini } 530628143c3dSStefano Zampini } 5307854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5308854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 530928143c3dSStefano Zampini for (i=1;i<nis;i++) { 531028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 531128143c3dSStefano Zampini } 531228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 531328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 531428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 531528143c3dSStefano Zampini for (j=0;j<nis;j++) { 531628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 531728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 531828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 531928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 532028143c3dSStefano Zampini } 532128143c3dSStefano Zampini } 532228143c3dSStefano Zampini for (i=0;i<nis;i++) { 532328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 532428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 532528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 532628143c3dSStefano Zampini } 532728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 532828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 532928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 533028143c3dSStefano Zampini } 5331e7931f94SStefano Zampini /* free workspace */ 533228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5333e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5334e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5335e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5336e7931f94SStefano Zampini if (isdense) { 5337e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5338e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5339e7931f94SStefano Zampini } else { 5340e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5341e7931f94SStefano Zampini } 534228143c3dSStefano Zampini if (nis) { 534328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 534428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 534528143c3dSStefano Zampini } 5346e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5347e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 534828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5349e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5350e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 535128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5352e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5353e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5354e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5355e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5356e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 535728143c3dSStefano Zampini if (nis) { 535828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 535928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 536028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 536128143c3dSStefano Zampini } 536228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 536328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 536428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 536528143c3dSStefano Zampini for (i=0;i<nis;i++) { 536628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 536728143c3dSStefano Zampini } 536853a05cb3SStefano Zampini *mat_n = NULL; 536928143c3dSStefano Zampini } 5370e7931f94SStefano Zampini PetscFunctionReturn(0); 5371e7931f94SStefano Zampini } 5372a57a6d2fSStefano Zampini 537312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5374af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 537512edc857SStefano Zampini 5376c8587f34SStefano Zampini #undef __FUNCT__ 5377c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5378c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5379c8587f34SStefano Zampini { 5380c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5381c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 538220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 53839881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 538420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 53856e683305SStefano Zampini IS coarse_is,*isarray; 53866e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 538730368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5388f9eb5b7dSStefano Zampini PC pc_temp; 5389c8587f34SStefano Zampini PCType coarse_pc_type; 5390c8587f34SStefano Zampini KSPType coarse_ksp_type; 5391f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 53924f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 53936e683305SStefano Zampini Mat t_coarse_mat_is; 53946e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 53956e683305SStefano Zampini PetscMPIInt all_procs; 539674e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 539768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 539822bc73bbSStefano Zampini PetscScalar *array; 53999881197aSStefano Zampini PetscErrorCode ierr; 5400fdc09c96SStefano Zampini 5401c8587f34SStefano Zampini PetscFunctionBegin; 5402c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 540368457ee5SStefano 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 */ 5404fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 54055a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5406fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5407f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5408f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5409f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5410fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 541151bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 541251bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5413dc4bcba2SStefano Zampini PC pc; 5414dc4bcba2SStefano Zampini PetscBool isbddc; 5415dc4bcba2SStefano Zampini 5416dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5417dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5418dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5419dc4bcba2SStefano Zampini if (isbddc) { 542063c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 542163c961adSStefano Zampini } else { 5422727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 542363c961adSStefano Zampini } 5424fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5425fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5426fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5427f4ddd8eeSStefano Zampini } 5428fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5429fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5430f4ddd8eeSStefano Zampini } 543170cf5478SStefano Zampini /* reset any subassembling information */ 543270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 54336e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54346e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5435fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5436f4ddd8eeSStefano Zampini } 5437c8587f34SStefano Zampini 54386e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 54392b510759SStefano Zampini im_active = !!(pcis->n); 5440b2566f29SBarry Smith ierr = MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 54416e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 54426e683305SStefano Zampini void_procs = all_procs-active_procs; 54436e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 544474e2c79eSStefano Zampini redist = PETSC_FALSE; 544522bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 54466e683305SStefano Zampini csin_ml = PETSC_TRUE; 54476e683305SStefano Zampini ncoarse_ml = void_procs; 5448779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5449779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 54506e683305SStefano Zampini csin_ds = PETSC_TRUE; 545118a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 545218a45a71SStefano Zampini redist = PETSC_TRUE; 545318a45a71SStefano Zampini } else { 54546e683305SStefano Zampini csin_ds = PETSC_TRUE; 5455779c1cceSStefano Zampini ncoarse_ds = active_procs; 5456779c1cceSStefano Zampini redist = PETSC_TRUE; 545718a45a71SStefano Zampini } 54586e683305SStefano Zampini } else { 54596e683305SStefano Zampini csin_ml = PETSC_FALSE; 54606e683305SStefano Zampini ncoarse_ml = all_procs; 54616e683305SStefano Zampini if (void_procs) { 5462*b9b0e38cSStefano Zampini csin_ml = PETSC_TRUE; 5463*b9b0e38cSStefano Zampini ncoarse_ml = void_procs; 54646e683305SStefano Zampini csin_ds = PETSC_TRUE; 54656e683305SStefano Zampini ncoarse_ds = void_procs; 54666e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 54676e683305SStefano Zampini } else { 5468779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 546974e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 547074e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 547174e2c79eSStefano Zampini redist = PETSC_TRUE; 547274e2c79eSStefano Zampini } else { 54736e683305SStefano Zampini csin_ds = PETSC_FALSE; 54746e683305SStefano Zampini ncoarse_ds = all_procs; 54756e683305SStefano Zampini } 54766e683305SStefano Zampini } 547774e2c79eSStefano Zampini } 54786e683305SStefano Zampini 54796e683305SStefano Zampini /* 54806e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 54816e683305SStefano Zampini - we have not exceeded the number of levels requested 54826e683305SStefano Zampini - we can actually subassemble the active processes 54836e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 54846e683305SStefano Zampini */ 54856e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 54866e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 54876e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 54886e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 54896e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5490f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 54912b510759SStefano Zampini } else { 5492f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5493c8587f34SStefano Zampini } 5494c8587f34SStefano Zampini } 54956e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 54966e683305SStefano Zampini if (multilevel_allowed) { 54976e683305SStefano Zampini ncoarse = ncoarse_ml; 54986e683305SStefano Zampini csin = csin_ml; 549958da7f69SStefano Zampini redist = PETSC_FALSE; 5500*b9b0e38cSStefano Zampini csin_type_simple = PETSC_TRUE; 55016e683305SStefano Zampini } else { 55026e683305SStefano Zampini ncoarse = ncoarse_ds; 55036e683305SStefano Zampini csin = csin_ds; 55046e683305SStefano Zampini } 5505e7931f94SStefano Zampini 5506abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5507abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5508abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5509abbbba34SStefano Zampini 5510abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 551122bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 551222bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 551322bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 551422bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5515e176bc59SStefano 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); 55166e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 55176e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 55186e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5519abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5520abbbba34SStefano Zampini 55216e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 552230368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 55236e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 55246e683305SStefano Zampini const PetscInt *idxs; 55256e683305SStefano Zampini ISLocalToGlobalMapping tmap; 55266e683305SStefano Zampini 55276e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 55280be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 55296e683305SStefano Zampini /* allocate space for temporary storage */ 5530854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5531854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 55326e683305SStefano Zampini /* allocate for IS array */ 55336e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 55346e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 553530368db7SStefano Zampini nisvert = 0; 553630368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 553730368db7SStefano Zampini nisvert = 1; 553830368db7SStefano Zampini } 553930368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5540854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 55416e683305SStefano Zampini /* dofs splitting */ 55426e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 55436e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 55446e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 55456e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55466e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55476e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 55486e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 554930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 55506e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 55516e683305SStefano Zampini } 55526e683305SStefano Zampini /* neumann boundaries */ 55536e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 55546e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 55556e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 55566e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55576e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 55586e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 55596e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 556030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 55616e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 55626e683305SStefano Zampini } 556330368db7SStefano Zampini /* primal vertices (benign) */ 556430368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 556530368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 556630368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 556730368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 556830368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 556930368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 557030368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 5571*b9b0e38cSStefano Zampini /* ierr = ISView(isarray[nis-1],0);CHKERRQ(ierr); */ 557230368db7SStefano Zampini } 55736e683305SStefano Zampini /* free memory */ 55746e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 55756e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 55766e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 55776e683305SStefano Zampini } else { 55786e683305SStefano Zampini nis = 0; 55796e683305SStefano Zampini nisdofs = 0; 55806e683305SStefano Zampini nisneu = 0; 558130368db7SStefano Zampini nisvert = 0; 55826e683305SStefano Zampini isarray = NULL; 55836e683305SStefano Zampini } 55846e683305SStefano Zampini /* destroy no longer needed map */ 55856e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 55866e683305SStefano Zampini 55876e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 55886e683305SStefano Zampini coarse_mat_is = NULL; 55896e683305SStefano Zampini if (csin) { 55906e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 559174e2c79eSStefano Zampini if (redist) { 559274e2c79eSStefano Zampini PetscMPIInt rank; 5593779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 559474e2c79eSStefano Zampini 559574e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 559658da7f69SStefano Zampini spc = active_procs/ncoarse; 559758da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5598779c1cceSStefano Zampini if (im_active) { 5599779c1cceSStefano Zampini destsize = 1; 560074e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 560174e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 560274e2c79eSStefano Zampini } else { 560374e2c79eSStefano Zampini dest[0] = rank/(spc+1); 560474e2c79eSStefano Zampini } 560574e2c79eSStefano Zampini } else { 5606779c1cceSStefano Zampini destsize = 0; 56076e683305SStefano Zampini } 5608779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5609779c1cceSStefano Zampini } else if (csin_type_simple) { 56106e683305SStefano Zampini PetscMPIInt rank; 56116e683305SStefano Zampini PetscInt issize,isidx; 5612779c1cceSStefano Zampini 56136e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 56146e683305SStefano Zampini if (im_active) { 56156e683305SStefano Zampini issize = 1; 56166e683305SStefano Zampini isidx = (PetscInt)rank; 56176e683305SStefano Zampini } else { 56186e683305SStefano Zampini issize = 0; 56196e683305SStefano Zampini isidx = -1; 56206e683305SStefano Zampini } 56216e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5622779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5623b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 56246e683305SStefano Zampini } 5625779c1cceSStefano Zampini 5626779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5627779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5628779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5629779c1cceSStefano Zampini PetscInt *coarse_candidates; 5630779c1cceSStefano Zampini const PetscInt* tisindices; 5631779c1cceSStefano Zampini 5632779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5633779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5634779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5635779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5636779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5637779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5638779c1cceSStefano Zampini } 5639779c1cceSStefano Zampini } 5640779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5641779c1cceSStefano Zampini 5642779c1cceSStefano Zampini 56436e683305SStefano Zampini if (pcbddc->dbg_flag) { 56446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56456e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 56466e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5648779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 56496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 56506e683305SStefano Zampini } 56516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 56526e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56536e683305SStefano Zampini } 56546e683305SStefano Zampini /* shift the pattern on coarse candidates */ 56556e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 56566e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5657854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 56586e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 56596e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 56606e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 56616e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 56626e683305SStefano Zampini } 56636e683305SStefano Zampini if (pcbddc->dbg_flag) { 56646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 56666e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 56676e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56686e683305SStefano Zampini } 5669779c1cceSStefano Zampini } 56706e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 567153a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 567253a05cb3SStefano 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); 567353a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 567453a05cb3SStefano 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); 567553a05cb3SStefano Zampini } 56766e683305SStefano Zampini } else { 56776e683305SStefano Zampini if (pcbddc->dbg_flag) { 56786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 56806e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56816e683305SStefano Zampini } 56826e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 56836e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 56846e683305SStefano Zampini } 56856e683305SStefano Zampini 56866e683305SStefano Zampini /* create local to global scatters for coarse problem */ 568768457ee5SStefano Zampini if (compute_vecs) { 56886e683305SStefano Zampini PetscInt lrows; 56896e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 56906e683305SStefano Zampini if (coarse_mat_is) { 56916e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 56926e683305SStefano Zampini } else { 56936e683305SStefano Zampini lrows = 0; 56946e683305SStefano Zampini } 56956e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 56966e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 56976e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 56986e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56996e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 57006e683305SStefano Zampini } 57016e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 57026e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5703c8587f34SStefano Zampini 5704f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5705f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5706f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5707f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5708f9eb5b7dSStefano Zampini } else { 5709f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5710f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5711c8587f34SStefano Zampini } 5712c8587f34SStefano Zampini 57136e683305SStefano Zampini /* print some info if requested */ 57146e683305SStefano Zampini if (pcbddc->dbg_flag) { 57156e683305SStefano Zampini if (!multilevel_allowed) { 57166e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57176e683305SStefano Zampini if (multilevel_requested) { 57186e683305SStefano 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); 57196e683305SStefano Zampini } else if (pcbddc->max_levels) { 57206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 57216e683305SStefano Zampini } 57226e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 57236e683305SStefano Zampini } 57246e683305SStefano Zampini } 57256e683305SStefano Zampini 5726f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 57276e683305SStefano Zampini if (coarse_mat_is) { 57286e683305SStefano Zampini MatReuse coarse_mat_reuse; 57296a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 57306e683305SStefano Zampini if (pcbddc->dbg_flag) { 57316e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 57326e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57336e683305SStefano Zampini } 5734f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5735312be037SStefano Zampini char prefix[256],str_level[16]; 5736e604994aSStefano Zampini size_t len; 57376e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5738422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5739c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5740f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 57415f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5742c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 57436e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5744c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5745c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5746e604994aSStefano Zampini /* prefix */ 5747e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5748e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5749e604994aSStefano Zampini if (!pcbddc->current_level) { 5750e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5751e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5752c8587f34SStefano Zampini } else { 5753e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5754312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5755312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 575634d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5757312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5758e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5759e604994aSStefano Zampini } 5760e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 57613e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 57623e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 57633e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 57643e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5765f9eb5b7dSStefano Zampini /* allow user customization */ 5766f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 57673e3c6dadSStefano Zampini } 57683e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 576951bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 57703e3c6dadSStefano Zampini if (nisdofs) { 57713e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 57723e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 57733e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 57743e3c6dadSStefano Zampini } 57753e3c6dadSStefano Zampini } 57763e3c6dadSStefano Zampini if (nisneu) { 57773e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 57783e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5779312be037SStefano Zampini } 578030368db7SStefano Zampini if (nisvert) { 578130368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 578230368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 578330368db7SStefano Zampini } 5784f9eb5b7dSStefano Zampini 5785f9eb5b7dSStefano Zampini /* get some info after set from options */ 5786f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5787f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 57884f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 57896e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5790f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5791f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5792f9eb5b7dSStefano Zampini } 579339f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 57944f3a063dSStefano Zampini if (isredundant) { 57954f3a063dSStefano Zampini KSP inner_ksp; 57964f3a063dSStefano Zampini PC inner_pc; 57974f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 57984f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 57994f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 58004f3a063dSStefano Zampini } 5801f9eb5b7dSStefano Zampini 5802f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5803fa7f1dd8SStefano Zampini if (coarse_reuse) { 580481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5805fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 58066e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5807fa7f1dd8SStefano Zampini } else { 58086e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5809fa7f1dd8SStefano Zampini } 5810c8587f34SStefano Zampini if (isbddc || isnn) { 5811720d30f9SStefano Zampini if (isbddc) { /* currently there are no APIs for these options */ 5812720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5813720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 5814720d30f9SStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_saddle_point; 5815720d30f9SStefano Zampini pcbddc_coarse->benign_compute_nonetflux = pcbddc->benign_compute_nonetflux; 5816720d30f9SStefano Zampini if (pcbddc_coarse->benign_compute_nonetflux) { 5817720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 5818720d30f9SStefano Zampini } 5819d4d8cf7bSStefano Zampini } 582022bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 582170cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5822b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 582322b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 58246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 58256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 58266e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 58276e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 582822b6e8a2SStefano Zampini } 582970cf5478SStefano Zampini } 583053a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 583170cf5478SStefano Zampini } else { 583222bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 583322bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 583422bc73bbSStefano Zampini } 583522bc73bbSStefano Zampini } else { 58362e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5837c8587f34SStefano Zampini } 5838c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5839c8587f34SStefano Zampini 58403301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 58415a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 58423301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 58433301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 58443301b35fSStefano Zampini } 58453301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 58463301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 58473301b35fSStefano Zampini } 58483301b35fSStefano Zampini if (pc->pmat->spd_set) { 58493301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 58503301b35fSStefano Zampini } 58516e683305SStefano Zampini /* set operators */ 58525f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 58536e683305SStefano Zampini if (pcbddc->dbg_flag) { 58546e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 58556e683305SStefano Zampini } 58566e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 58576e683305SStefano Zampini coarse_mat = 0; 58586e683305SStefano Zampini } 58596e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5860b1ecc7b1SStefano Zampini #if 0 5861b9b85e73SStefano Zampini { 5862b9b85e73SStefano Zampini PetscViewer viewer; 5863b9b85e73SStefano Zampini char filename[256]; 5864b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5865b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 58666a9046bcSBarry Smith ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5867b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5868f159cad9SBarry Smith ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 5869b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5870b9b85e73SStefano Zampini } 5871b9b85e73SStefano Zampini #endif 5872c8587f34SStefano Zampini 5873c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5874298c0119SStefano Zampini #if 0 5875c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5876c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 587798a51de6SStefano Zampini } 5878298c0119SStefano Zampini #endif 5879b0f5fe93SStefano Zampini /* hack */ 588098a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 588198a51de6SStefano Zampini Vec crhs,csol; 588204708bb6SStefano Zampini 5883f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5884f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5885f347579bSStefano Zampini if (!csol) { 58862a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5887f9eb5b7dSStefano Zampini } 5888f347579bSStefano Zampini if (!crhs) { 58892a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5890f347579bSStefano Zampini } 5891b0f5fe93SStefano Zampini } 5892b0f5fe93SStefano Zampini 5893b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5894b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5895b0f5fe93SStefano Zampini 5896b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 58974f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58984f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 58994f1b2e48SStefano Zampini } 5900b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5901b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5902b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5903b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5904b0f5fe93SStefano Zampini if (coarse_mat) { 5905b0f5fe93SStefano Zampini Vec nullv; 5906b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5907b0f5fe93SStefano Zampini PetscInt nl; 5908b0f5fe93SStefano Zampini 5909b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5910b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5911b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5912b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5913b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5914b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5915b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5916b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5917b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5918b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5919b0f5fe93SStefano Zampini } 5920b0f5fe93SStefano Zampini } 5921b0f5fe93SStefano Zampini 5922b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5923b0f5fe93SStefano Zampini PetscBool ispreonly; 5924b0f5fe93SStefano Zampini 5925b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5926b0f5fe93SStefano Zampini PetscBool isnull; 5927b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5928d4d8cf7bSStefano Zampini if (0) { 592930368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5930b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5931b0f5fe93SStefano Zampini } else { 5932b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5933b0f5fe93SStefano Zampini } 5934b0f5fe93SStefano Zampini } else { 5935b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5936b0f5fe93SStefano Zampini } 5937b0f5fe93SStefano Zampini } 5938b0f5fe93SStefano Zampini /* setup coarse ksp */ 5939b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5940cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5941cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 59426e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5943c8587f34SStefano Zampini KSP check_ksp; 59442b510759SStefano Zampini KSPType check_ksp_type; 5945c8587f34SStefano Zampini PC check_pc; 59466e683305SStefano Zampini Vec check_vec,coarse_vec; 59476a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 59482b510759SStefano Zampini PetscInt its; 59496e683305SStefano Zampini PetscBool compute_eigs; 59506e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 59516e683305SStefano Zampini PetscInt neigs; 59528e185a42SStefano Zampini const char *prefix; 5953c8587f34SStefano Zampini 59542b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 59556e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5956422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 595723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5958f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 59592b510759SStefano Zampini if (ispreonly) { 59602b510759SStefano Zampini check_ksp_type = KSPPREONLY; 59616e683305SStefano Zampini compute_eigs = PETSC_FALSE; 59622b510759SStefano Zampini } else { 5963cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 59646e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5965c8587f34SStefano Zampini } 5966c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 59676e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 59686e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 59696e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5970a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5971a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5972a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5973a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5974c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5975c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5976c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5977c8587f34SStefano Zampini /* create random vec */ 59786e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 59796e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5980c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5981c8587f34SStefano Zampini if (CoarseNullSpace) { 5982c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5983c8587f34SStefano Zampini } 59846e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5985c8587f34SStefano Zampini /* solve coarse problem */ 59866e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5987c8587f34SStefano Zampini if (CoarseNullSpace) { 59886e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5989c8587f34SStefano Zampini } 5990cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 59916e683305SStefano Zampini if (compute_eigs) { 5992854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5993854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 59946e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 59956e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 59966e683305SStefano Zampini lambda_min = eigs_r[0]; 59976e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 59986e683305SStefano Zampini if (lambda_max>lambda_min) { 5999cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 6000cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 6001cbcc2c2aSStefano Zampini } 6002c8587f34SStefano Zampini } 6003c8587f34SStefano Zampini } 6004cbcc2c2aSStefano Zampini 6005c8587f34SStefano Zampini /* check coarse problem residual error */ 60066e683305SStefano Zampini if (pcbddc->dbg_flag) { 60076e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 60086e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 60096e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 6010c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 60116e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 60126e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 6013c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 6014779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 60156e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 60166e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 60176e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 60186e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 6019b0f5fe93SStefano Zampini if (CoarseNullSpace) { 6020b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 6021b0f5fe93SStefano Zampini } 60226e683305SStefano Zampini if (compute_eigs) { 60236e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 6024deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 6025c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 60266e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 60276e683305SStefano 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); 60286e683305SStefano Zampini for (i=0;i<neigs;i++) { 60296e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 6030c8587f34SStefano Zampini } 60316e683305SStefano Zampini } 60326e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 60336e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 60346e683305SStefano Zampini } 6035c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 60366e683305SStefano Zampini if (compute_eigs) { 60376e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 60386e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 6039c8587f34SStefano Zampini } 60406e683305SStefano Zampini } 60416e683305SStefano Zampini } 6042cbcc2c2aSStefano Zampini /* print additional info */ 6043cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 60446e683305SStefano Zampini /* waits until all processes reaches this point */ 60456e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 6046cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 6047cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6048cbcc2c2aSStefano Zampini } 6049cbcc2c2aSStefano Zampini 60502b510759SStefano Zampini /* free memory */ 6051c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 6052fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 6053c8587f34SStefano Zampini PetscFunctionReturn(0); 6054c8587f34SStefano Zampini } 6055674ae819SStefano Zampini 6056f34684f1SStefano Zampini #undef __FUNCT__ 6057f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 6058f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 6059f34684f1SStefano Zampini { 6060f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 6061f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 6062f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 6063dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 6064dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 606573be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 6066dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 6067f34684f1SStefano Zampini PetscErrorCode ierr; 6068f34684f1SStefano Zampini 6069f34684f1SStefano Zampini PetscFunctionBegin; 6070f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 60716c4ed002SBarry Smith if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 6072dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 60733bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 6074dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6075dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6076dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6077dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6078dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6079dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60806c4ed002SBarry 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); 6081dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6082dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6083dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6084dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6085dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6086f34684f1SStefano Zampini 6087f34684f1SStefano Zampini /* check numbering */ 6088f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6089019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6090dc456d91SStefano Zampini PetscInt i; 6091b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6092f34684f1SStefano Zampini 6093f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6094f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6095f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 60961575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6097019a44ceSStefano Zampini /* counter */ 6098019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6099019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6100019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6101019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6102019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6103019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6104f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6105f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6106727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6107f34684f1SStefano Zampini } 6108f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6109f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6110f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6111e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6112e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6113e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6114e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6115f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6116019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6117f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6118019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 611975c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 612075c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6121b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6122019a44ceSStefano 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); 6123f34684f1SStefano Zampini } 6124f34684f1SStefano Zampini } 6125019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6126b2566f29SBarry Smith ierr = MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6127f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6128f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6129f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6130f34684f1SStefano Zampini } 6131f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6132f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6133e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6134e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6135f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6136f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6137b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6138ca8b9ea9SStefano Zampini PetscInt *gidxs; 6139ca8b9ea9SStefano Zampini 6140ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 61413bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6142f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6143f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6144f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6145f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 61464bc2dc4bSStefano 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); 6147f34684f1SStefano Zampini } 6148f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6149ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6150f34684f1SStefano Zampini } 6151f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 61521575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6153302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6154f34684f1SStefano Zampini } 61558bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6156f34684f1SStefano Zampini /* get back data */ 6157f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6158f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6159674ae819SStefano Zampini PetscFunctionReturn(0); 6160674ae819SStefano Zampini } 6161674ae819SStefano Zampini 6162e456f2a8SStefano Zampini #undef __FUNCT__ 6163e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6164a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6165e456f2a8SStefano Zampini { 6166e456f2a8SStefano Zampini IS localis_t; 6167a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6168e456f2a8SStefano Zampini PetscScalar *vals; 6169e456f2a8SStefano Zampini PetscErrorCode ierr; 6170e456f2a8SStefano Zampini 6171e456f2a8SStefano Zampini PetscFunctionBegin; 6172a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6173e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6174854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6175e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6176e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6177a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6178a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61791035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6180a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61811035eff8SStefano Zampini } 6182a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6183e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6184e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6185a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6186a7dc3881SStefano Zampini /* now compute set in local ordering */ 6187a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6188a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6189a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6190a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6191a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6192ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6193e456f2a8SStefano Zampini lsize++; 6194e456f2a8SStefano Zampini } 6195e456f2a8SStefano Zampini } 6196854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6197a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6198ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6199e456f2a8SStefano Zampini idxs[lsize++] = i; 6200e456f2a8SStefano Zampini } 6201e456f2a8SStefano Zampini } 6202a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6203a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6204e456f2a8SStefano Zampini *localis = localis_t; 6205e456f2a8SStefano Zampini PetscFunctionReturn(0); 6206e456f2a8SStefano Zampini } 6207906d46d4SStefano Zampini 6208b96c3477SStefano Zampini #undef __FUNCT__ 6209b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 621008122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6211b96c3477SStefano Zampini { 6212a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6213b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6214b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6215a64f4aa4SStefano Zampini Mat S_j; 6216b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6217b96c3477SStefano Zampini PetscBool free_used_adj; 6218b96c3477SStefano Zampini PetscErrorCode ierr; 6219b96c3477SStefano Zampini 6220b96c3477SStefano Zampini PetscFunctionBegin; 6221b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6222b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 622308122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6224b96c3477SStefano Zampini used_xadj = NULL; 6225b96c3477SStefano Zampini used_adjncy = NULL; 6226b96c3477SStefano Zampini } else { 622708122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 622808122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 622908122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 623008122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6231b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6232b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6233b96c3477SStefano Zampini } else { 62342fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6235b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6236b96c3477SStefano Zampini PetscInt nvtxs; 6237b96c3477SStefano Zampini 62382fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 62392fffb893SStefano Zampini if (flg_row) { 6240b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6241b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6242b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6243b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62442fffb893SStefano Zampini } else { 62452fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62462fffb893SStefano Zampini used_xadj = NULL; 62472fffb893SStefano Zampini used_adjncy = NULL; 62482fffb893SStefano Zampini } 62492fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6250b96c3477SStefano Zampini } 6251b96c3477SStefano Zampini } 6252d5574798SStefano Zampini 6253d5574798SStefano Zampini /* setup sub_schurs data */ 6254a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6255df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6256df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6257a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 62589d54b7f4SStefano 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); 6259a64f4aa4SStefano Zampini } else { 62606816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 6261b7ab4a40SStefano Zampini PetscBool isseqaij,need_change = PETSC_FALSE;; 6262a3df083aSStefano Zampini PetscInt benign_n; 626372b8c272SStefano Zampini Mat change = NULL; 62649d54b7f4SStefano Zampini Vec scaling = NULL; 626572b8c272SStefano Zampini IS change_primal = NULL; 6266a3df083aSStefano Zampini 62675feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62685feab87aSStefano Zampini PetscInt n_vertices; 62695feab87aSStefano Zampini 62705feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62712034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62725feab87aSStefano Zampini } 627304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 627404708bb6SStefano Zampini if (!isseqaij) { 627504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 627604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 627704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 627804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 627904708bb6SStefano Zampini } else { 6280511c6705SHong Zhang ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 628104708bb6SStefano Zampini } 628204708bb6SStefano Zampini } 6283a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6284a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6285ca92afb2SStefano Zampini } else { 6286a3df083aSStefano Zampini benign_n = 0; 6287ca92afb2SStefano Zampini } 6288b7ab4a40SStefano Zampini /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc. 6289b7ab4a40SStefano Zampini We need a global reduction to avoid possible deadlocks. 6290b7ab4a40SStefano Zampini We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */ 629172b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 6292b7ab4a40SStefano Zampini PetscBool have_loc_change = !!(sub_schurs->change); 6293b7ab4a40SStefano Zampini ierr = MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6294b7ab4a40SStefano Zampini need_change = !need_change; 6295b7ab4a40SStefano Zampini } 6296b7ab4a40SStefano Zampini /* If the user defines additional constraints, we import them here. 6297b7ab4a40SStefano 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 */ 6298b7ab4a40SStefano Zampini if (need_change) { 629988c03ad3SStefano Zampini PC_IS *pcisf; 630088c03ad3SStefano Zampini PC_BDDC *pcbddcf; 630188c03ad3SStefano Zampini PC pcf; 630288c03ad3SStefano Zampini 630388c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 630488c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 630588c03ad3SStefano Zampini } 630688c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 630788c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 630888c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 630988c03ad3SStefano Zampini /* hacks */ 631088c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 631172b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 631272b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 631372b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 631472b8c272SStefano Zampini pcisf->n = pcis->n; 631572b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 631688c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 631788c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 631888c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 631988c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 632088c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 632188c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 632272b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 632388c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 632488c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 632572b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 632672b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 632772b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 632872b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 632972b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 633088c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 633172b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 633288c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 633388c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 633488c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 633588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 633688c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 633788c03ad3SStefano Zampini pcf->ops->destroy = NULL; 633888c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 633988c03ad3SStefano Zampini } 63409d54b7f4SStefano Zampini if (!pcbddc->use_deluxe_scaling) scaling = pcis->D; 63419d54b7f4SStefano 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); 634272b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 634372b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6344ca92afb2SStefano Zampini } 6345d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6346b96c3477SStefano Zampini 6347b96c3477SStefano Zampini /* free adjacency */ 6348b96c3477SStefano Zampini if (free_used_adj) { 6349b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6350b96c3477SStefano Zampini } 6351b96c3477SStefano Zampini PetscFunctionReturn(0); 6352b96c3477SStefano Zampini } 6353b96c3477SStefano Zampini 6354b96c3477SStefano Zampini #undef __FUNCT__ 6355b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 635608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6357b96c3477SStefano Zampini { 6358b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6359b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6360b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6361b96c3477SStefano Zampini PCBDDCGraph graph; 6362b96c3477SStefano Zampini PetscErrorCode ierr; 6363b96c3477SStefano Zampini 6364b96c3477SStefano Zampini PetscFunctionBegin; 6365b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 636608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63673301b35fSStefano Zampini IS verticesIS,verticescomm; 63683301b35fSStefano Zampini PetscInt vsize,*idxs; 6369b96c3477SStefano Zampini 6370b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63713301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63723301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63733301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63743301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63753301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6376b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63777fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 6378441e0de0SStefano Zampini ierr = PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63793301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6380b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6381b96c3477SStefano Zampini /* 6382b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6383b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6384b96c3477SStefano Zampini } 6385b96c3477SStefano Zampini */ 6386b96c3477SStefano Zampini } else { 6387b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6388b96c3477SStefano Zampini } 6389b96c3477SStefano Zampini 6390b96c3477SStefano Zampini /* sub_schurs init */ 6391a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6392a64f4aa4SStefano Zampini 6393b96c3477SStefano Zampini /* free graph struct */ 639408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6395b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6396b96c3477SStefano Zampini } 6397b96c3477SStefano Zampini PetscFunctionReturn(0); 6398b96c3477SStefano Zampini } 6399fa34dd3eSStefano Zampini 6400fa34dd3eSStefano Zampini #undef __FUNCT__ 6401fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6402fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6403fa34dd3eSStefano Zampini { 6404fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6405fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6406fa34dd3eSStefano Zampini PetscErrorCode ierr; 6407fa34dd3eSStefano Zampini 6408fa34dd3eSStefano Zampini PetscFunctionBegin; 6409fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6410fa34dd3eSStefano Zampini IS zerodiag = NULL; 64114f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6412fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 64134f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 641475c01103SStefano Zampini PetscReal norm; 6415fa34dd3eSStefano Zampini PetscInt i; 6416fa34dd3eSStefano Zampini 6417fa34dd3eSStefano Zampini /* B0 and B0_B */ 6418fa34dd3eSStefano Zampini if (zerodiag) { 6419fa34dd3eSStefano Zampini IS dummy; 6420fa34dd3eSStefano Zampini 64214f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 64224f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6423fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6424fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6425fa34dd3eSStefano Zampini } 6426fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6427fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6428fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6429fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6430fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6431fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6432fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6433fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6434fa34dd3eSStefano Zampini /* S_j */ 6435fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6436fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6437fa34dd3eSStefano Zampini 6438fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6439fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6440fa34dd3eSStefano Zampini /* continuous in primal space */ 6441fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6442fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6443fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6444fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 64454f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 64464f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6447fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6448fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6449fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6450fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6451fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6452fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6453fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6454fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6455fa34dd3eSStefano Zampini 6456fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6457fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6458fa34dd3eSStefano Zampini /* local with Schur */ 6459fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6460fa34dd3eSStefano Zampini if (zerodiag) { 6461fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64624f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6463fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6464fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6465fa34dd3eSStefano Zampini } 6466fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6467fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6468fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6469fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6470fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6471fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6472fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6473fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6474fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6475fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6476fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6477fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6478fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6479fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6480fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6481fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6482fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6483fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6484fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6485fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6486fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6487fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6488fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6489fa34dd3eSStefano Zampini if (zerodiag) { 6490fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6491fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64924f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6493fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6494fa34dd3eSStefano Zampini } 6495fa34dd3eSStefano Zampini /* BDDC */ 6496fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6497fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6498fa34dd3eSStefano Zampini 6499fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6500fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6501fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6502fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 65034f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 65044f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6505fa34dd3eSStefano Zampini } 65064f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6507fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6508fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6509fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6510fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6511fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6512fa34dd3eSStefano Zampini } 6513fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6514fa34dd3eSStefano Zampini } 6515