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 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 8906d46d4SStefano Zampini 9a3df083aSStefano Zampini #undef __FUNCT__ 10a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 11a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 12a3df083aSStefano Zampini { 13a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 14a3df083aSStefano Zampini PetscErrorCode ierr; 15a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 16a3df083aSStefano Zampini 17a3df083aSStefano Zampini PetscFunctionBegin; 18a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 19a3df083aSStefano Zampini if (transpose) { 20a3df083aSStefano Zampini apply_right = ctx->apply_left; 21a3df083aSStefano Zampini apply_left = ctx->apply_right; 22a3df083aSStefano Zampini } else { 23a3df083aSStefano Zampini apply_right = ctx->apply_right; 24a3df083aSStefano Zampini apply_left = ctx->apply_left; 25a3df083aSStefano Zampini } 26a3df083aSStefano Zampini reset_x = PETSC_FALSE; 27a3df083aSStefano Zampini if (apply_right) { 28a3df083aSStefano Zampini const PetscScalar *ax; 29a3df083aSStefano Zampini PetscInt nl,i; 30a3df083aSStefano Zampini 31a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 32a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 33a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 34a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 35a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 36a3df083aSStefano Zampini PetscScalar sum,val; 37a3df083aSStefano Zampini const PetscInt *idxs; 38a3df083aSStefano Zampini PetscInt nz,j; 39a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 40a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 41a3df083aSStefano Zampini sum = 0.; 42a3df083aSStefano Zampini if (ctx->apply_p0) { 43a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 44a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 45a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 46a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 47a3df083aSStefano Zampini } 48a3df083aSStefano Zampini } else { 49a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 50a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 51a3df083aSStefano Zampini } 52a3df083aSStefano Zampini } 53a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 54a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 55a3df083aSStefano Zampini } 56a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 57a3df083aSStefano Zampini reset_x = PETSC_TRUE; 58a3df083aSStefano Zampini } 59a3df083aSStefano Zampini if (transpose) { 60a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 61a3df083aSStefano Zampini } else { 62a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 63a3df083aSStefano Zampini } 64a3df083aSStefano Zampini if (reset_x) { 65a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 66a3df083aSStefano Zampini } 67a3df083aSStefano Zampini if (apply_left) { 68a3df083aSStefano Zampini PetscScalar *ay; 69a3df083aSStefano Zampini PetscInt i; 70a3df083aSStefano Zampini 71a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 72a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 73a3df083aSStefano Zampini PetscScalar sum,val; 74a3df083aSStefano Zampini const PetscInt *idxs; 75a3df083aSStefano Zampini PetscInt nz,j; 76a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 77a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 78a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 79a3df083aSStefano Zampini if (ctx->apply_p0) { 80a3df083aSStefano Zampini sum = 0.; 81a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 82a3df083aSStefano Zampini sum += ay[idxs[j]]; 83a3df083aSStefano Zampini ay[idxs[j]] += val; 84a3df083aSStefano Zampini } 85a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 86a3df083aSStefano Zampini } else { 87a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 88a3df083aSStefano Zampini ay[idxs[j]] += val; 89a3df083aSStefano Zampini } 90a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 91a3df083aSStefano Zampini } 92a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 93a3df083aSStefano Zampini } 94a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 95a3df083aSStefano Zampini } 96a3df083aSStefano Zampini PetscFunctionReturn(0); 97a3df083aSStefano Zampini } 98a3df083aSStefano Zampini 99a3df083aSStefano Zampini #undef __FUNCT__ 100a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 101a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 102a3df083aSStefano Zampini { 103a3df083aSStefano Zampini PetscErrorCode ierr; 104a3df083aSStefano Zampini 105a3df083aSStefano Zampini PetscFunctionBegin; 106a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 107a3df083aSStefano Zampini PetscFunctionReturn(0); 108a3df083aSStefano Zampini } 109a3df083aSStefano Zampini 110a3df083aSStefano Zampini #undef __FUNCT__ 111a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 112a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 113a3df083aSStefano Zampini { 114a3df083aSStefano Zampini PetscErrorCode ierr; 115a3df083aSStefano Zampini 116a3df083aSStefano Zampini PetscFunctionBegin; 117a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 118a3df083aSStefano Zampini PetscFunctionReturn(0); 119a3df083aSStefano Zampini } 120a3df083aSStefano Zampini 121a3df083aSStefano Zampini #undef __FUNCT__ 122a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 123a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 124a3df083aSStefano Zampini { 125a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 126a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 127a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 128a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 129a3df083aSStefano Zampini PetscErrorCode ierr; 130a3df083aSStefano Zampini 131a3df083aSStefano Zampini PetscFunctionBegin; 132a3df083aSStefano Zampini if (!restore) { 133a3df083aSStefano Zampini Mat A,A_IB,A_BI; 134a3df083aSStefano Zampini PetscScalar *work; 135df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 136a3df083aSStefano Zampini 137a3df083aSStefano Zampini if (!reuse) { 138a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot setup shell matrices!"); 139a3df083aSStefano Zampini } 140a3df083aSStefano Zampini 141a3df083aSStefano Zampini /* local mat */ 142a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = MatSetSizes(A,pcis->n,pcis->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatSetType(A,MATSHELL);CHKERRQ(ierr); 145a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 146a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 147a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 148a3df083aSStefano Zampini ierr = MatShellSetContext(A,ctx);CHKERRQ(ierr); 149a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 150a3df083aSStefano Zampini ctx->apply_right = PETSC_TRUE; 151a3df083aSStefano Zampini ctx->apply_p0 = PETSC_TRUE; 152a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 153a3df083aSStefano Zampini ctx->benign_zerodiag_subs = pcbddc->benign_zerodiag_subs; 154a3df083aSStefano Zampini ctx->A = matis->A; 155a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 156a3df083aSStefano Zampini ctx->work = work; 157a3df083aSStefano Zampini ierr = MatSetUp(A);CHKERRQ(ierr); 158a3df083aSStefano Zampini ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 159a3df083aSStefano Zampini ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 160a3df083aSStefano Zampini matis->A = A; 161a3df083aSStefano Zampini 162a3df083aSStefano Zampini /* A_IB */ 163a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 164a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 165a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 166a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 167a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 168a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 169a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 170a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 171a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 172a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 173a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 174a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 175a3df083aSStefano Zampini ctx->A = pcis->A_IB; 176a3df083aSStefano Zampini ctx->work = work; 177a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 178a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 179a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 180a3df083aSStefano Zampini pcis->A_IB = A_IB; 181a3df083aSStefano Zampini 182a3df083aSStefano Zampini /* A_BI as A_IB^T */ 183a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 184a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 185a3df083aSStefano Zampini pcis->A_BI = A_BI; 186a3df083aSStefano Zampini } else { 187a3df083aSStefano Zampini ierr = MatShellGetContext(matis->A,&ctx);CHKERRQ(ierr); 188a3df083aSStefano Zampini ierr = MatDestroy(&matis->A);CHKERRQ(ierr); 189a3df083aSStefano Zampini matis->A = ctx->A; 190a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 191a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 192a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 193a3df083aSStefano Zampini pcis->A_IB = ctx->A; 194a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 195a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 196a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 197a3df083aSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 198a3df083aSStefano Zampini pcbddc->benign_original_mat = NULL; 199a3df083aSStefano Zampini } 200a3df083aSStefano Zampini PetscFunctionReturn(0); 201a3df083aSStefano Zampini } 202a3df083aSStefano Zampini 203a3df083aSStefano Zampini /* used just in bddc debug mode */ 204a3df083aSStefano Zampini #undef __FUNCT__ 205a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 206a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 207a3df083aSStefano Zampini { 208a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 209a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 210a3df083aSStefano Zampini Mat An; 211a3df083aSStefano Zampini PetscErrorCode ierr; 212a3df083aSStefano Zampini 213a3df083aSStefano Zampini PetscFunctionBegin; 214a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 215a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 216a3df083aSStefano Zampini if (is1) { 217a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 218a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 219a3df083aSStefano Zampini } else { 220a3df083aSStefano Zampini *B = An; 221a3df083aSStefano Zampini } 222a3df083aSStefano Zampini PetscFunctionReturn(0); 223a3df083aSStefano Zampini } 224a3df083aSStefano Zampini 2251cf9b237SStefano Zampini /* TODO: add reuse flag */ 2261cf9b237SStefano Zampini #undef __FUNCT__ 2271cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2281cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2291cf9b237SStefano Zampini { 2301cf9b237SStefano Zampini Mat Bt; 2311cf9b237SStefano Zampini PetscScalar *a,*bdata; 2321cf9b237SStefano Zampini const PetscInt *ii,*ij; 2331cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2341cf9b237SStefano Zampini PetscBool flg_row; 2351cf9b237SStefano Zampini PetscErrorCode ierr; 2361cf9b237SStefano Zampini 2371cf9b237SStefano Zampini PetscFunctionBegin; 2381cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2391cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2401cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2411cf9b237SStefano Zampini nnz = n; 2421cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2431cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2441cf9b237SStefano Zampini } 2451cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2461cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2471cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2481cf9b237SStefano Zampini nnz = 0; 2491cf9b237SStefano Zampini bii[0] = 0; 2501cf9b237SStefano Zampini for (i=0;i<n;i++) { 2511cf9b237SStefano Zampini PetscInt j; 2521cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2531cf9b237SStefano Zampini PetscScalar entry = a[j]; 2541cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2551cf9b237SStefano Zampini bij[nnz] = ij[j]; 2561cf9b237SStefano Zampini bdata[nnz] = entry; 2571cf9b237SStefano Zampini nnz++; 2581cf9b237SStefano Zampini } 2591cf9b237SStefano Zampini } 2601cf9b237SStefano Zampini bii[i+1] = nnz; 2611cf9b237SStefano Zampini } 2621cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2631cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2641cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2651cf9b237SStefano Zampini { 2661cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2671cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2681cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2691cf9b237SStefano Zampini } 2701cf9b237SStefano Zampini *B = Bt; 2711cf9b237SStefano Zampini PetscFunctionReturn(0); 2721cf9b237SStefano Zampini } 2731cf9b237SStefano Zampini 274674ae819SStefano Zampini #undef __FUNCT__ 2754f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2764f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2774f1b2e48SStefano Zampini { 2784f1b2e48SStefano Zampini Mat B; 2794f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2804f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2814f1b2e48SStefano Zampini PCBDDCGraph graph; 2824f1b2e48SStefano Zampini PetscInt i,n; 2834f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 2844f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 2854f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 2864f1b2e48SStefano Zampini PetscErrorCode ierr; 2874f1b2e48SStefano Zampini 2884f1b2e48SStefano Zampini PetscFunctionBegin; 28963c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 29063c961adSStefano Zampini *ncc = 0; 29163c961adSStefano Zampini *cc = NULL; 29263c961adSStefano Zampini PetscFunctionReturn(0); 29363c961adSStefano Zampini } 2944f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 2954f1b2e48SStefano Zampini if (!isseqaij && filter) { 2961cf9b237SStefano Zampini PetscBool isseqdense; 2971cf9b237SStefano Zampini 2981cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 2991cf9b237SStefano Zampini if (!isseqdense) { 3004f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3011cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3021cf9b237SStefano Zampini PetscScalar *array; 3031cf9b237SStefano Zampini PetscReal chop=1.e-6; 3041cf9b237SStefano Zampini 3051cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3061cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3071cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3081cf9b237SStefano Zampini for (i=0;i<n;i++) { 3091cf9b237SStefano Zampini PetscInt j; 3101cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3111cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3121cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3131cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3141cf9b237SStefano Zampini } 3151cf9b237SStefano Zampini } 3161cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3171cf9b237SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); 3181cf9b237SStefano Zampini } 3194f1b2e48SStefano Zampini } else { 3204f1b2e48SStefano Zampini B = A; 3214f1b2e48SStefano Zampini } 3224f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3234f1b2e48SStefano Zampini 3244f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3254f1b2e48SStefano Zampini if (filter) { 3264f1b2e48SStefano Zampini PetscScalar *data; 3274f1b2e48SStefano Zampini PetscInt j,cum; 3284f1b2e48SStefano Zampini 3294f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3304f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3314f1b2e48SStefano Zampini cum = 0; 3324f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3334f1b2e48SStefano Zampini PetscInt t; 3344f1b2e48SStefano Zampini 3354f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3364f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3374f1b2e48SStefano Zampini continue; 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3404f1b2e48SStefano Zampini } 3414f1b2e48SStefano Zampini t = xadj_filtered[i]; 3424f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3434f1b2e48SStefano Zampini cum += t; 3444f1b2e48SStefano Zampini } 3454f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3464f1b2e48SStefano Zampini } else { 3474f1b2e48SStefano Zampini xadj_filtered = NULL; 3484f1b2e48SStefano Zampini adjncy_filtered = NULL; 3494f1b2e48SStefano Zampini } 3504f1b2e48SStefano Zampini 3514f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3524f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3534f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3544f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3554f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3564f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3574f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3584f1b2e48SStefano Zampini if (xadj_filtered) { 3594f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3604f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3614f1b2e48SStefano Zampini } else { 3624f1b2e48SStefano Zampini graph->xadj = xadj; 3634f1b2e48SStefano Zampini graph->adjncy = adjncy; 3644f1b2e48SStefano Zampini } 3654f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3664f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3674f1b2e48SStefano Zampini /* partial clean up */ 3684f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3694f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3701cf9b237SStefano Zampini if (A != B) { 3714f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3724f1b2e48SStefano Zampini } 3734f1b2e48SStefano Zampini 3744f1b2e48SStefano Zampini /* get back data */ 3751cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3761cf9b237SStefano Zampini if (cc) { 3774f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3784f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3794f1b2e48SStefano 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); 3804f1b2e48SStefano Zampini } 3814f1b2e48SStefano Zampini *cc = cc_n; 3821cf9b237SStefano Zampini } 3834f1b2e48SStefano Zampini /* clean up graph */ 3844f1b2e48SStefano Zampini graph->xadj = 0; 3854f1b2e48SStefano Zampini graph->adjncy = 0; 3864f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 3874f1b2e48SStefano Zampini PetscFunctionReturn(0); 3884f1b2e48SStefano Zampini } 3894f1b2e48SStefano Zampini 3904f1b2e48SStefano Zampini #undef __FUNCT__ 3915408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 3925408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 3935408967cSStefano Zampini { 3945408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3955408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 396dee84bffSStefano Zampini IS dirIS = NULL; 3974f1b2e48SStefano Zampini PetscInt i; 3985408967cSStefano Zampini PetscErrorCode ierr; 3995408967cSStefano Zampini 4005408967cSStefano Zampini PetscFunctionBegin; 401dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4025408967cSStefano Zampini if (zerodiag) { 4035408967cSStefano Zampini Mat A; 4045408967cSStefano Zampini Vec vec3_N; 4055408967cSStefano Zampini PetscScalar *vals; 4065408967cSStefano Zampini const PetscInt *idxs; 407d12d3064SStefano Zampini PetscInt nz,*count; 4085408967cSStefano Zampini 4095408967cSStefano Zampini /* p0 */ 4105408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4115408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4125408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4135408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4144f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4155408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4165408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4175408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4185408967cSStefano Zampini /* v_I */ 4195408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4205408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4215408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4225408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4235408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4245408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4255408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4265408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4275408967cSStefano Zampini if (dirIS) { 4285408967cSStefano Zampini PetscInt n; 4295408967cSStefano Zampini 4305408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4315408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4325408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4335408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4345408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4355408967cSStefano Zampini } 4365408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4375408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4385408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4415408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4425408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 4434f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 4445408967cSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %f (should be numerically 0.)",PetscAbsScalar(vals[0])); 4455408967cSStefano Zampini } 4465408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4475408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 448d12d3064SStefano Zampini 449d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 450d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 451d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 452d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 453d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 454d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 455d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 456d12d3064SStefano Zampini if (count[idxs[i]]) { 457d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 458d12d3064SStefano Zampini } 459d4d8cf7bSStefano Zampini } 460d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 461d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4625408967cSStefano Zampini } 463dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4645408967cSStefano Zampini 4655408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4665408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4674f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4685408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4694f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4705408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4714f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4724f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4734f1b2e48SStefano 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); 4744f1b2e48SStefano Zampini } 4755408967cSStefano Zampini } 4765408967cSStefano Zampini PetscFunctionReturn(0); 4775408967cSStefano Zampini } 4785408967cSStefano Zampini 4795408967cSStefano Zampini #undef __FUNCT__ 480339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 481339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 482339f8db1SStefano Zampini { 483339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4844f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 485b0f5fe93SStefano Zampini PetscInt nz,n; 4864f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 487339f8db1SStefano Zampini PetscErrorCode ierr; 488339f8db1SStefano Zampini 489339f8db1SStefano Zampini PetscFunctionBegin; 4909f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4919f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 492339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 493a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 494a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 495a3df083aSStefano Zampini } 496a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 497a3df083aSStefano Zampini pcbddc->benign_n = 0; 4984f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4994f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 5004f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 5014f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 5024f1b2e48SStefano Zampini since the local Schur complements are SPD 5034f1b2e48SStefano Zampini */ 5044f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5054f1b2e48SStefano Zampini have_null = PETSC_TRUE; 50640fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5074f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5084f1b2e48SStefano Zampini 5094f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5104f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5114f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5124f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 513ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 51440fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 51540fa8d13SStefano Zampini if (!sorted) { 51640fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 51740fa8d13SStefano Zampini } 51840fa8d13SStefano Zampini } else { 51940fa8d13SStefano Zampini pressures = NULL; 52040fa8d13SStefano Zampini } 52197d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 52297d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 52397d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 524339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 525339f8db1SStefano Zampini if (!sorted) { 526339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 527339f8db1SStefano Zampini } 528339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5294f1b2e48SStefano Zampini if (!nz) { 5304f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5314f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 53240fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 53340fa8d13SStefano Zampini } 5344f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5354f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5364f1b2e48SStefano Zampini zerodiag_subs = NULL; 5374f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5384f1b2e48SStefano Zampini if (has_null_pressures) { 5394f1b2e48SStefano Zampini IS *subs; 5404f1b2e48SStefano Zampini PetscInt nsubs,i; 5414f1b2e48SStefano Zampini 5424f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5434f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5444f1b2e48SStefano Zampini if (nsubs > 1) { 5454f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5464f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5474f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5484f1b2e48SStefano Zampini IS t_zerodiag_subs; 5494f1b2e48SStefano Zampini PetscInt nl; 5504f1b2e48SStefano Zampini 5514f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5524f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5534f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5544f1b2e48SStefano Zampini if (nl) { 5554f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5564f1b2e48SStefano Zampini 5574f1b2e48SStefano Zampini if (pressures) { 5584f1b2e48SStefano Zampini IS t_pressure_subs; 5594f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5604f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5614f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5624f1b2e48SStefano Zampini } 5634f1b2e48SStefano Zampini if (valid) { 5644f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5654f1b2e48SStefano Zampini pcbddc->benign_n++; 5664f1b2e48SStefano Zampini } else { 5674f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5684f1b2e48SStefano Zampini } 5694f1b2e48SStefano Zampini } 5704f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5714f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5724f1b2e48SStefano Zampini } 5734f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5744f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5754f1b2e48SStefano Zampini if (pressures) { 5764f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5774f1b2e48SStefano Zampini } 5784f1b2e48SStefano Zampini if (valid) { 5794f1b2e48SStefano Zampini pcbddc->benign_n = 1; 580ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5814f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5824f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5834f1b2e48SStefano Zampini } 5844f1b2e48SStefano Zampini } 5854f1b2e48SStefano Zampini } 5864f1b2e48SStefano Zampini 5874f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 5884f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5894f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5904f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5914f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5924f1b2e48SStefano Zampini } 5934f1b2e48SStefano Zampini 5944f1b2e48SStefano Zampini /* final check for null pressures */ 5954f1b2e48SStefano Zampini if (zerodiag && pressures) { 5964f1b2e48SStefano Zampini PetscInt nz,np; 5974f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5984f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 5994f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 6004f1b2e48SStefano Zampini } 6014f1b2e48SStefano Zampini 6024f1b2e48SStefano Zampini if (recompute_zerodiag) { 6034f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6044f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6054f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6074f1b2e48SStefano Zampini } else { 6084f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6094f1b2e48SStefano Zampini 6104f1b2e48SStefano Zampini nzn = 0; 6114f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6124f1b2e48SStefano Zampini PetscInt ns; 6134f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6144f1b2e48SStefano Zampini nzn += ns; 6154f1b2e48SStefano Zampini } 6164f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6174f1b2e48SStefano Zampini nzn = 0; 6184f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6194f1b2e48SStefano Zampini PetscInt ns,*idxs; 6204f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6224f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6234f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6244f1b2e48SStefano Zampini nzn += ns; 6254f1b2e48SStefano Zampini } 6264f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6274f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6284f1b2e48SStefano Zampini } 6294f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6304f1b2e48SStefano Zampini } 6314f1b2e48SStefano Zampini 6324f1b2e48SStefano Zampini if (has_null_pressures) { 6334f1b2e48SStefano Zampini IS zerodiagc; 6344f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 6354f1b2e48SStefano Zampini PetscInt i,s,*nnz; 6364f1b2e48SStefano Zampini 6374f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 638339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 639339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 640339f8db1SStefano Zampini /* local change of basis for pressures */ 641339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 64297d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 643339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 644339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 645339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 6464f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 6474f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6484f1b2e48SStefano Zampini PetscInt nzs,j; 6494f1b2e48SStefano Zampini 6504f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 6514f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6524f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 6534f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 6544f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6554f1b2e48SStefano Zampini } 656339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 657339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 658339f8db1SStefano Zampini /* set identity on velocities */ 659339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 660339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 661339f8db1SStefano Zampini } 6624f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 6634f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 6649f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 6654f1b2e48SStefano 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); 666339f8db1SStefano Zampini /* set change on pressures */ 6674f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 6684f1b2e48SStefano Zampini PetscScalar *array; 6694f1b2e48SStefano Zampini PetscInt nzs; 6704f1b2e48SStefano Zampini 6714f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 6724f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 6734f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 674339f8db1SStefano Zampini PetscScalar vals[2]; 675339f8db1SStefano Zampini PetscInt cols[2]; 676339f8db1SStefano Zampini 677339f8db1SStefano Zampini cols[0] = idxs[i]; 6784f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 679339f8db1SStefano Zampini vals[0] = 1.; 680b0f5fe93SStefano Zampini vals[1] = 1.; 6814f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 682339f8db1SStefano Zampini } 6834f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 6844f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 6854f1b2e48SStefano Zampini array[nzs-1] = 1.; 6864f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 6874f1b2e48SStefano Zampini /* store local idxs for p0 */ 6884f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 6894f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 690339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 6914f1b2e48SStefano Zampini } 692339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 693339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 694a3df083aSStefano Zampini /* project if needed */ 695a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 69697d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 697339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 69897d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 69997d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 700a3df083aSStefano Zampini } 7014f1b2e48SStefano Zampini /* store global idxs for p0 */ 7024f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 703339f8db1SStefano Zampini } 704ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7054f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 706b0f5fe93SStefano Zampini 707b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 708b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 709339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 710339f8db1SStefano Zampini PetscFunctionReturn(0); 711339f8db1SStefano Zampini } 712339f8db1SStefano Zampini 713339f8db1SStefano Zampini #undef __FUNCT__ 714015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 715015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 716efc2fbd9SStefano Zampini { 717efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 718efc2fbd9SStefano Zampini PetscErrorCode ierr; 719efc2fbd9SStefano Zampini 720efc2fbd9SStefano Zampini PetscFunctionBegin; 721efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 722efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 7234f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 724efc2fbd9SStefano Zampini } 725015636ebSStefano Zampini if (get) { /* use SF to get values */ 726efc2fbd9SStefano Zampini PetscScalar *array; 727efc2fbd9SStefano Zampini 728efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7294f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 7304f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 731efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7324f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 7334f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 734efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 735efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 736efc2fbd9SStefano Zampini } 737efc2fbd9SStefano Zampini PetscFunctionReturn(0); 738efc2fbd9SStefano Zampini } 739efc2fbd9SStefano Zampini 740efc2fbd9SStefano Zampini #undef __FUNCT__ 741c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 742c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 743c263805aSStefano Zampini { 744c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 745c263805aSStefano Zampini PetscErrorCode ierr; 746c263805aSStefano Zampini 747c263805aSStefano Zampini PetscFunctionBegin; 748c263805aSStefano Zampini /* TODO: add error checking 749c263805aSStefano Zampini - avoid nested pop (or push) calls. 750c263805aSStefano Zampini - cannot push before pop. 7511c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 752c263805aSStefano Zampini */ 7534f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 754efc2fbd9SStefano Zampini PetscFunctionReturn(0); 755efc2fbd9SStefano Zampini } 756c263805aSStefano Zampini if (pop) { 757a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7584f1b2e48SStefano Zampini IS is_p0; 7594f1b2e48SStefano Zampini MatReuse reuse; 760c263805aSStefano Zampini 761c263805aSStefano Zampini /* extract B_0 */ 7624f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 7634f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 7644f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 7654f1b2e48SStefano Zampini } 7664f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 7674f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 768c263805aSStefano Zampini /* remove rows and cols from local problem */ 769c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 77097d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 7714f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 7724f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 773a3df083aSStefano Zampini } else { 774a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 775a3df083aSStefano Zampini PetscScalar *vals; 776a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 777a3df083aSStefano Zampini 778a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 779a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 780a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 781a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 782a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 783a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 784*331e053bSStefano Zampini if (!pcbddc->current_level) { 785*331e053bSStefano Zampini /* this matrix is actually sparse: the nnz pattern is not known unless we do 2 sweeps of the next loop. 786*331e053bSStefano Zampini Setting nnz=100 should be more than enough for the finest level */ 787a3df083aSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,100,NULL);CHKERRQ(ierr); 788*331e053bSStefano Zampini } else { 789*331e053bSStefano Zampini PetscInt *nnz; 790*331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 791*331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 792*331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 793*331e053bSStefano Zampini nnz[i] = n - nnz[i]; 794*331e053bSStefano Zampini } 795*331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 796*331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 797*331e053bSStefano Zampini } 798a3df083aSStefano Zampini } 799a3df083aSStefano Zampini 800a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 801a3df083aSStefano Zampini PetscScalar *array; 802a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 803a3df083aSStefano Zampini 804a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 805a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 806a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 807a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 808a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 809a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 810a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 811a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 812a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 813a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 814a3df083aSStefano Zampini cum = 0; 815a3df083aSStefano Zampini for (j=0;j<n;j++) { 816a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 817a3df083aSStefano Zampini vals[cum] = array[j]; 818a3df083aSStefano Zampini idxs_ins[cum] = j; 819a3df083aSStefano Zampini cum++; 820a3df083aSStefano Zampini } 821a3df083aSStefano Zampini } 822a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 823a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 824a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 825a3df083aSStefano Zampini } 826a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 827a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 828a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 829a3df083aSStefano Zampini } 830c263805aSStefano Zampini } else { /* push */ 831a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8324f1b2e48SStefano Zampini PetscInt i; 8334f1b2e48SStefano Zampini 8344f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8354f1b2e48SStefano Zampini PetscScalar *B0_vals; 8364f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 8374f1b2e48SStefano Zampini 8384f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8394f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8407b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8414f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 8424f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8434f1b2e48SStefano Zampini } 844c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 845c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 846a3df083aSStefano Zampini } else { 847a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 848a3df083aSStefano Zampini } 849c263805aSStefano Zampini } 850c263805aSStefano Zampini PetscFunctionReturn(0); 851c263805aSStefano Zampini } 852c263805aSStefano Zampini 853c263805aSStefano Zampini #undef __FUNCT__ 854b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 85508122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 856b1b3d7a2SStefano Zampini { 857b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 85808122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 85908122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 86008122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 86108122e43SStefano Zampini PetscScalar *work,lwork; 86208122e43SStefano Zampini PetscScalar *St,*S,*eigv; 86308122e43SStefano Zampini PetscScalar *Sarray,*Starray; 86408122e43SStefano Zampini PetscReal *eigs,thresh; 8651b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 866f6f667cfSStefano Zampini PetscBool allocated_S_St; 86708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 86808122e43SStefano Zampini PetscReal *rwork; 86908122e43SStefano Zampini #endif 870b1b3d7a2SStefano Zampini PetscErrorCode ierr; 871b1b3d7a2SStefano Zampini 872b1b3d7a2SStefano Zampini PetscFunctionBegin; 873df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 874df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 87508122e43SStefano Zampini } 87608122e43SStefano Zampini 87706a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 87806a4e24aSStefano Zampini 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); 87906a4e24aSStefano Zampini } 88006a4e24aSStefano Zampini 881fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 882fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 883fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 884fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 8851575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 886fd14bc51SStefano Zampini } 887fd14bc51SStefano Zampini 888e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 889e496cd5dSStefano 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); 890e496cd5dSStefano Zampini } 891e496cd5dSStefano Zampini 89208122e43SStefano Zampini /* max size of subsets */ 89308122e43SStefano Zampini mss = 0; 89408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 89508122e43SStefano Zampini PetscInt subset_size; 896862806e4SStefano Zampini 89708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 89808122e43SStefano Zampini mss = PetscMax(mss,subset_size); 89908122e43SStefano Zampini } 90008122e43SStefano Zampini 90108122e43SStefano Zampini /* min/max and threshold */ 90208122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 903f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 90408122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 905f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 906f6f667cfSStefano Zampini if (nmin) { 907f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 908f6f667cfSStefano Zampini } 90908122e43SStefano Zampini 91008122e43SStefano Zampini /* allocate lapack workspace */ 91108122e43SStefano Zampini cum = cum2 = 0; 91208122e43SStefano Zampini maxneigs = 0; 91308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 91408122e43SStefano Zampini PetscInt n,subset_size; 915f6f667cfSStefano Zampini 91608122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 91708122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9189162d606SStefano Zampini cum += subset_size; 9199162d606SStefano Zampini cum2 += subset_size*n; 92008122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 92108122e43SStefano Zampini } 92208122e43SStefano Zampini if (mss) { 9239ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 92408122e43SStefano Zampini PetscBLASInt B_itype = 1; 92508122e43SStefano Zampini PetscBLASInt B_N = mss; 9264c6709b3SStefano Zampini PetscReal zero = 0.0; 9274c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 92808122e43SStefano Zampini 92908122e43SStefano Zampini B_lwork = -1; 93008122e43SStefano Zampini S = NULL; 93108122e43SStefano Zampini St = NULL; 932a58a30b4SStefano Zampini eigs = NULL; 933a58a30b4SStefano Zampini eigv = NULL; 934a58a30b4SStefano Zampini B_iwork = NULL; 935a58a30b4SStefano Zampini B_ifail = NULL; 936d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 937d1710679SStefano Zampini rwork = NULL; 938d1710679SStefano Zampini #endif 9398bec7fa6SStefano Zampini thresh = 1.0; 94008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 94108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 94208122e43SStefano 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)); 94308122e43SStefano Zampini #else 94408122e43SStefano 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)); 94508122e43SStefano Zampini #endif 94608122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 94708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 94808122e43SStefano Zampini } else { 94908122e43SStefano Zampini /* TODO */ 95008122e43SStefano Zampini } 95108122e43SStefano Zampini } else { 95208122e43SStefano Zampini lwork = 0; 95308122e43SStefano Zampini } 95408122e43SStefano Zampini 95508122e43SStefano Zampini nv = 0; 956d62866d3SStefano 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) */ 957d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 95808122e43SStefano Zampini } 9594c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 960f6f667cfSStefano Zampini if (allocated_S_St) { 961f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 962f6f667cfSStefano Zampini } 963f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 96408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 96508122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 96608122e43SStefano Zampini #endif 9679162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 9689162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 9699162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 97008122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 9719162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 97208122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 97308122e43SStefano Zampini 97408122e43SStefano Zampini maxneigs = 0; 97508122e43SStefano Zampini cum = cum2 = cumarray = 0; 9769162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 9779162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 978d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 97908122e43SStefano Zampini const PetscInt *idxs; 98008122e43SStefano Zampini 981d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 98208122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 98308122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 98408122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 98508122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 9869162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 9879162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 98808122e43SStefano Zampini } 98908122e43SStefano Zampini cum2 = cum; 990d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 99108122e43SStefano Zampini } 99208122e43SStefano Zampini 99308122e43SStefano Zampini if (mss) { /* multilevel */ 99408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 99508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 99608122e43SStefano Zampini } 99708122e43SStefano Zampini 998ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 99908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 100008122e43SStefano Zampini const PetscInt *idxs; 1001f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1002862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 100308122e43SStefano Zampini PetscBLASInt B_N; 1004aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 100508122e43SStefano Zampini 1006862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1007ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1008f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1009f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10109ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1011aff50787SStefano Zampini PetscInt j,k; 1012aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1013aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1014aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 101508122e43SStefano Zampini } 101608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1017aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1018aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1019aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1020aff50787SStefano Zampini } 102108122e43SStefano Zampini } 102208122e43SStefano Zampini } else { 102308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 102408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 102508122e43SStefano Zampini } 10268bec7fa6SStefano Zampini } else { 1027f6f667cfSStefano Zampini S = Sarray + cumarray; 1028f6f667cfSStefano Zampini St = Starray + cumarray; 10298bec7fa6SStefano Zampini } 103008122e43SStefano Zampini 1031aff50787SStefano Zampini /* see if we can save some work */ 1032aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1033aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1034aff50787SStefano Zampini } 1035aff50787SStefano Zampini 1036aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1037aff50787SStefano Zampini B_neigs = 0; 1038aff50787SStefano Zampini } else { 10399ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 104008122e43SStefano Zampini PetscBLASInt B_itype = 1; 1041f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 10424c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 10439552c7c7SStefano Zampini PetscInt nmin_s; 104408122e43SStefano Zampini 1045fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 10468bec7fa6SStefano 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]]); 1047fd14bc51SStefano Zampini } 1048d16cbb6bSStefano Zampini 104908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1050d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1051d16cbb6bSStefano Zampini 1052d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 105308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1054f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 105508122e43SStefano Zampini #else 1056f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 105708122e43SStefano Zampini #endif 1058d16cbb6bSStefano Zampini } else { 1059d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1060d16cbb6bSStefano Zampini B_IL = 1; 1061d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1062d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1063d16cbb6bSStefano Zampini #else 1064d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1065d16cbb6bSStefano Zampini #endif 1066d16cbb6bSStefano Zampini } 106708122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 106808122e43SStefano Zampini if (B_ierr) { 106908122e43SStefano Zampini if (B_ierr < 0 ) { 107008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 107108122e43SStefano Zampini } else if (B_ierr <= B_N) { 107208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 107308122e43SStefano Zampini } else { 10749552c7c7SStefano Zampini 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); 107508122e43SStefano Zampini } 107608122e43SStefano Zampini } 107708122e43SStefano Zampini 107808122e43SStefano Zampini if (B_neigs > nmax) { 1079fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1080fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1081fd14bc51SStefano Zampini } 1082f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 108308122e43SStefano Zampini B_neigs = nmax; 108408122e43SStefano Zampini } 108508122e43SStefano Zampini 10869552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 10879552c7c7SStefano Zampini if (B_neigs < nmin_s) { 108808122e43SStefano Zampini PetscBLASInt B_neigs2; 108908122e43SStefano Zampini 1090f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1091f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1092fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1093fd14bc51SStefano 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); 1094fd14bc51SStefano Zampini } 10959ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 109608122e43SStefano Zampini PetscInt j; 109708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 109808122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 109908122e43SStefano Zampini } 110008122e43SStefano Zampini for (j=0;j<subset_size;j++) { 110108122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 110208122e43SStefano Zampini } 110308122e43SStefano Zampini } else { 110408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110608122e43SStefano Zampini } 110708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 110808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1109f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&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)); 111008122e43SStefano Zampini #else 1111f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&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)); 111208122e43SStefano Zampini #endif 111308122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 111408122e43SStefano Zampini B_neigs += B_neigs2; 111508122e43SStefano Zampini } 111608122e43SStefano Zampini if (B_ierr) { 111708122e43SStefano Zampini if (B_ierr < 0 ) { 111808122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 111908122e43SStefano Zampini } else if (B_ierr <= B_N) { 112008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 112108122e43SStefano Zampini } else { 11229552c7c7SStefano Zampini 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); 112308122e43SStefano Zampini } 112408122e43SStefano Zampini } 1125fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1126ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 112708122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 112808122e43SStefano Zampini if (eigs[j] == 0.0) { 1129ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 113008122e43SStefano Zampini } else { 1131ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1132fd14bc51SStefano Zampini } 113308122e43SStefano Zampini } 113408122e43SStefano Zampini } 113508122e43SStefano Zampini } else { 113608122e43SStefano Zampini /* TODO */ 113708122e43SStefano Zampini } 1138aff50787SStefano Zampini } 11398bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 11408bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 11419162d606SStefano Zampini if (B_neigs) { 11429162d606SStefano 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); 1143fd14bc51SStefano Zampini 1144fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 11459552c7c7SStefano Zampini PetscInt ii; 11469552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1147ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11489552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1149ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1150ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1151ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1152ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1153ac47001eSStefano Zampini #else 1154ac47001eSStefano 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); 1155ac47001eSStefano Zampini #endif 11569552c7c7SStefano Zampini } 11579552c7c7SStefano Zampini } 1158fd14bc51SStefano Zampini } 115908122e43SStefano Zampini #if 0 11609162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 116108122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 116208122e43SStefano Zampini PetscScalar norm; 116308122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 11649162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 11659162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 116608122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 116708122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 116808122e43SStefano Zampini } else { 116908122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 117008122e43SStefano Zampini } 11719162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 1172b1b3d7a2SStefano Zampini } 1173b1b3d7a2SStefano Zampini #endif 11749162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 11759162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 11769162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 11779162d606SStefano Zampini cum++; 117808122e43SStefano Zampini } 117908122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 118008122e43SStefano Zampini /* shift for next computation */ 118108122e43SStefano Zampini cumarray += subset_size*subset_size; 118208122e43SStefano Zampini } 1183fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1184fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1185fd14bc51SStefano Zampini } 118608122e43SStefano Zampini 118708122e43SStefano Zampini if (mss) { 118808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 118908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1190f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1191f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1192f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 119308122e43SStefano Zampini } 1194f6f667cfSStefano Zampini if (allocated_S_St) { 1195f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1196f6f667cfSStefano Zampini } 1197f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 119808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 119908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 120008122e43SStefano Zampini #endif 120108122e43SStefano Zampini if (pcbddc->dbg_flag) { 12021b968477SStefano Zampini PetscInt maxneigs_r; 120308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12049b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 120508122e43SStefano Zampini } 120608122e43SStefano Zampini PetscFunctionReturn(0); 120708122e43SStefano Zampini } 1208b1b3d7a2SStefano Zampini 1209674ae819SStefano Zampini #undef __FUNCT__ 1210c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1211c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1212c8587f34SStefano Zampini { 1213c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12148629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1215c8587f34SStefano Zampini PetscErrorCode ierr; 1216c8587f34SStefano Zampini 1217c8587f34SStefano Zampini PetscFunctionBegin; 1218f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12195e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1220c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1221c8587f34SStefano Zampini 1222684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12230fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1224684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1225c8587f34SStefano Zampini 1226c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1227b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1228c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1229c8587f34SStefano Zampini } 1230c8587f34SStefano Zampini 12318629588bSStefano Zampini /* 12328629588bSStefano Zampini Setup local correction and local part of coarse basis. 12338629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12348629588bSStefano Zampini */ 123547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12368629588bSStefano Zampini 12378629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12388629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12398629588bSStefano Zampini 12408629588bSStefano Zampini /* free */ 12418629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1242c8587f34SStefano Zampini PetscFunctionReturn(0); 1243c8587f34SStefano Zampini } 1244c8587f34SStefano Zampini 1245c8587f34SStefano Zampini #undef __FUNCT__ 1246674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1247674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1248674ae819SStefano Zampini { 1249674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1250674ae819SStefano Zampini PetscErrorCode ierr; 1251674ae819SStefano Zampini 1252674ae819SStefano Zampini PetscFunctionBegin; 1253674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1254674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 125530368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1256674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1257674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1258785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1259674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1260f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1261f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1262785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 126363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 126463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1265674ae819SStefano Zampini PetscFunctionReturn(0); 1266674ae819SStefano Zampini } 1267674ae819SStefano Zampini 1268674ae819SStefano Zampini #undef __FUNCT__ 1269674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1270674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1271674ae819SStefano Zampini { 1272674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12734f1b2e48SStefano Zampini PetscInt i; 1274674ae819SStefano Zampini PetscErrorCode ierr; 1275674ae819SStefano Zampini 1276674ae819SStefano Zampini PetscFunctionBegin; 1277b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1278674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1279674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1280674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 12814f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 12824f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 12834f1b2e48SStefano Zampini } 12844f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1285b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1286674ae819SStefano Zampini PetscFunctionReturn(0); 1287674ae819SStefano Zampini } 1288674ae819SStefano Zampini 1289674ae819SStefano Zampini #undef __FUNCT__ 1290674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1291674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1292674ae819SStefano Zampini { 1293674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1294674ae819SStefano Zampini PetscErrorCode ierr; 1295674ae819SStefano Zampini 1296674ae819SStefano Zampini PetscFunctionBegin; 1297674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 129858da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1299ca92afb2SStefano Zampini PetscScalar *array; 130006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 130106656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 130258da7f69SStefano Zampini } 1303674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1304674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 130515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 130615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1307674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1308674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1309674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 131006656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1311674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1312674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13138ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1314674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1315674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1316674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1317f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1318f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1319f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1320f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1321727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13220e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1323f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 132470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 13256e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 132681d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 13270369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 13288b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 13294f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13308b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1331ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1332ca92afb2SStefano Zampini PetscInt i; 1333ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1334ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1335ca92afb2SStefano Zampini } 1336ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1337ca92afb2SStefano Zampini } 13384f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1339674ae819SStefano Zampini PetscFunctionReturn(0); 1340674ae819SStefano Zampini } 1341674ae819SStefano Zampini 1342674ae819SStefano Zampini #undef __FUNCT__ 1343f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1344f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 13456bfb1811SStefano Zampini { 13466bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13476bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 13486bfb1811SStefano Zampini VecType impVecType; 13494f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 13506bfb1811SStefano Zampini PetscErrorCode ierr; 13516bfb1811SStefano Zampini 13526bfb1811SStefano Zampini PetscFunctionBegin; 1353f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1354019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1355f4ddd8eeSStefano Zampini } 1356e7b262bdSStefano Zampini /* get sizes */ 13574f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1358b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 13596bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1360e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1361e7b262bdSStefano Zampini /* R nodes */ 1362e7b262bdSStefano Zampini old_size = -1; 1363e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1364e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1365e7b262bdSStefano Zampini } 1366e7b262bdSStefano Zampini if (n_R != old_size) { 1367e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1368e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13696bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 13706bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 13716bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 13726bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1373e7b262bdSStefano Zampini } 1374e7b262bdSStefano Zampini /* local primal dofs */ 1375e7b262bdSStefano Zampini old_size = -1; 1376e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1377e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1378e7b262bdSStefano Zampini } 1379e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1380e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 138183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1382e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 13836bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1384e7b262bdSStefano Zampini } 1385e7b262bdSStefano Zampini /* local explicit constraints */ 1386e7b262bdSStefano Zampini old_size = -1; 1387e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1388e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1389e7b262bdSStefano Zampini } 1390e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1391e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 139283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 139383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 139483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 139583b7ccabSStefano Zampini } 13966bfb1811SStefano Zampini PetscFunctionReturn(0); 13976bfb1811SStefano Zampini } 13986bfb1811SStefano Zampini 13996bfb1811SStefano Zampini #undef __FUNCT__ 140047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 140147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 140288ebb749SStefano Zampini { 140325084f0cSStefano Zampini PetscErrorCode ierr; 140425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 140588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 140688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1407d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 140825084f0cSStefano Zampini /* submatrices of local problem */ 140980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1410019a44ceSStefano Zampini /* submatrices of benign trick */ 1411d16cbb6bSStefano Zampini Mat B0_V = NULL; 141206656605SStefano Zampini /* submatrices of local coarse problem */ 141306656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 141425084f0cSStefano Zampini /* working matrices */ 141506656605SStefano Zampini Mat C_CR; 141625084f0cSStefano Zampini /* additional working stuff */ 141706656605SStefano Zampini PC pc_R; 14184f1b2e48SStefano Zampini Mat F; 1419a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 142006656605SStefano Zampini 142125084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 142206656605SStefano Zampini PetscScalar *work; 142306656605SStefano Zampini PetscInt *idx_V_B; 1424ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 142506656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1426ffd830a3SStefano Zampini 142725084f0cSStefano Zampini /* some shortcuts to scalars */ 142806656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 142988ebb749SStefano Zampini 143088ebb749SStefano Zampini PetscFunctionBegin; 1431ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1432ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1433ffd830a3SStefano Zampini } 1434ffd830a3SStefano Zampini 1435ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1436b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14374f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1438b371cd4fSStefano Zampini n_B = pcis->n_B; 1439b371cd4fSStefano Zampini n_D = pcis->n - n_B; 144088ebb749SStefano Zampini n_R = pcis->n - n_vertices; 144188ebb749SStefano Zampini 144288ebb749SStefano Zampini /* vertices in boundary numbering */ 1443785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 14440e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 144588ebb749SStefano Zampini if (i != n_vertices) { 144622d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 144788ebb749SStefano Zampini } 144888ebb749SStefano Zampini 144906656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1450019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 145106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 145206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 145306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 145406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 145506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 145606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 145706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 145806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 145906656605SStefano Zampini 146006656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 146106656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 146206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 146306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 146406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1465ffd830a3SStefano Zampini lda_rhs = n_R; 1466a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 146706656605SStefano Zampini if (isLU || isILU || isCHOL) { 146806656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1469df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1470df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1471d62866d3SStefano Zampini MatFactorType type; 1472d62866d3SStefano Zampini 1473df4d28bfSStefano Zampini F = reuse_solver->F; 14746816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1475d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1476ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1477df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 147806656605SStefano Zampini } else { 147906656605SStefano Zampini F = NULL; 148006656605SStefano Zampini } 148106656605SStefano Zampini 1482ffd830a3SStefano Zampini /* allocate workspace */ 1483ffd830a3SStefano Zampini n = 0; 1484ffd830a3SStefano Zampini if (n_constraints) { 1485ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1486ffd830a3SStefano Zampini } 1487ffd830a3SStefano Zampini if (n_vertices) { 1488ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1489ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1490ffd830a3SStefano Zampini } 1491ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1492ffd830a3SStefano Zampini 149388ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 149488ebb749SStefano Zampini if (n_constraints) { 149506656605SStefano Zampini Mat M1,M2,M3; 149680677318SStefano Zampini Mat auxmat; 149706656605SStefano Zampini IS is_aux; 149880677318SStefano Zampini PetscScalar *array,*array2; 149906656605SStefano Zampini 1500f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 150180677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 150288ebb749SStefano Zampini 150325084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 150425084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15058ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 150680677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 150788ebb749SStefano Zampini 150880677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 150980677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1510ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 151188ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 151206656605SStefano Zampini const PetscScalar *row_cmat_values; 151306656605SStefano Zampini const PetscInt *row_cmat_indices; 151406656605SStefano Zampini PetscInt size_of_constraint,j; 151588ebb749SStefano Zampini 151606656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 151706656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1518ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 151906656605SStefano Zampini } 152006656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 152106656605SStefano Zampini } 1522ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 152306656605SStefano Zampini if (F) { 152406656605SStefano Zampini Mat B; 152506656605SStefano Zampini 1526ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1527a3df083aSStefano Zampini if (need_benign_correction) { 1528a3df083aSStefano Zampini PetscScalar *marr; 1529df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1530a3df083aSStefano Zampini 1531a3df083aSStefano Zampini ierr = MatDenseGetArray(B,&marr);CHKERRQ(ierr); 1532a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1533a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1534df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1535a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1536a3df083aSStefano Zampini } 1537a3df083aSStefano Zampini ierr = MatDenseRestoreArray(B,&marr);CHKERRQ(ierr); 1538a3df083aSStefano Zampini } 153980677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1540a3df083aSStefano Zampini if (need_benign_correction) { 1541a3df083aSStefano Zampini PetscScalar *marr; 1542df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1543a3df083aSStefano Zampini 1544a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1545a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1546a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1547df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1548a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1549a3df083aSStefano Zampini } 1550a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1551a3df083aSStefano Zampini } 155206656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 155306656605SStefano Zampini } else { 155480677318SStefano Zampini PetscScalar *marr; 155580677318SStefano Zampini 155680677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 155706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1558ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1559ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 156006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 156106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 156206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 156306656605SStefano Zampini } 156480677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 156506656605SStefano Zampini } 156680677318SStefano Zampini if (!pcbddc->switch_static) { 156780677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 156880677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 156980677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 157080677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1571ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 157280677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 157380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 157480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 157580677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 157680677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 157780677318SStefano Zampini } 157880677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 157980677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 158080677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 158180677318SStefano Zampini } else { 1582ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1583ffd830a3SStefano Zampini IS dummy; 1584ffd830a3SStefano Zampini 1585ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1586ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1587ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1588ffd830a3SStefano Zampini } else { 158980677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 159080677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1591ffd830a3SStefano Zampini } 159225084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 159380677318SStefano Zampini } 159480677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 159580677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 159680677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 159706656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 159806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 159980677318SStefano Zampini if (isCHOL) { 160080677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 160180677318SStefano Zampini } else { 160225084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 160380677318SStefano Zampini } 160480677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 160506656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 160625084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 160725084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 160825084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 160980677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 161080677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 161180677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 161206656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 161306656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1614f4ddd8eeSStefano Zampini } 161588ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 16164f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1617d16cbb6bSStefano Zampini IS dummy; 1618d16cbb6bSStefano Zampini Mat B0_R; 1619d16cbb6bSStefano Zampini PetscReal norm; 1620d16cbb6bSStefano Zampini 16214f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16224f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1623d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1624d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1625d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1626d16cbb6bSStefano Zampini } 1627d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1628d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1629d16cbb6bSStefano Zampini } 1630d16cbb6bSStefano Zampini 163188ebb749SStefano Zampini if (n_vertices) { 163206656605SStefano Zampini IS is_aux; 16333a50541eSStefano Zampini 1634df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16356816873aSStefano Zampini IS tis; 16366816873aSStefano Zampini 16376816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16386816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16396816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16406816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16416816873aSStefano Zampini } else { 16423a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16436816873aSStefano Zampini } 16449577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16459577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 164604708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 16474f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1648019a44ceSStefano Zampini IS dummy; 1649019a44ceSStefano Zampini 16504f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16514f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1652019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1653019a44ceSStefano Zampini } 165425084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 165588ebb749SStefano Zampini } 165688ebb749SStefano Zampini 165788ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1658f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 165906656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 166006656605SStefano Zampini if (pcbddc->coarse_phi_D) { 166106656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 166206656605SStefano Zampini } 1663f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 166406656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 166506656605SStefano Zampini PetscScalar *marray; 166606656605SStefano Zampini 166706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 166806656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1669f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1670f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1671f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1672f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1673f4ddd8eeSStefano Zampini } 1674f4ddd8eeSStefano Zampini } 167506656605SStefano Zampini 1676f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 167706656605SStefano Zampini PetscScalar *marray; 167888ebb749SStefano Zampini 167906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16808eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 168106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 168288ebb749SStefano Zampini } 16833301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 168406656605SStefano Zampini n *= 2; 168588ebb749SStefano Zampini } 168606656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 168706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 168806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16898eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 169006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 169106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 169288ebb749SStefano Zampini } 16933301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 169406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 16958eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 169606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 169706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 169888ebb749SStefano Zampini } 169988ebb749SStefano Zampini } else { 1700c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1701c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17021b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1703c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1704c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1705c0553b1fSStefano Zampini } 170688ebb749SStefano Zampini } 170706656605SStefano Zampini } 1708019a44ceSStefano Zampini 170906656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17104f1b2e48SStefano Zampini p0_lidx_I = NULL; 17114f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1712d12edf2fSStefano Zampini const PetscInt *idxs; 1713d12edf2fSStefano Zampini 1714d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17154f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17164f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17174f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17184f1b2e48SStefano Zampini } 1719d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1720d12edf2fSStefano Zampini } 1721d16cbb6bSStefano Zampini 172206656605SStefano Zampini /* vertices */ 172306656605SStefano Zampini if (n_vertices) { 172416f15bc4SStefano Zampini 1725ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 172604708bb6SStefano Zampini 172716f15bc4SStefano Zampini if (n_R) { 172814393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 172906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 173016f15bc4SStefano Zampini PetscScalar *x,*y; 173104708bb6SStefano Zampini PetscBool isseqaij; 173206656605SStefano Zampini 173321eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 173414393ed6SStefano Zampini if (need_benign_correction) { 173514393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 173614393ed6SStefano Zampini IS is_p0; 173714393ed6SStefano Zampini PetscInt *idxs_p0,n; 173814393ed6SStefano Zampini 173914393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 174014393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 174114393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 174214393ed6SStefano Zampini if (n != pcbddc->benign_n) { 174314393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 174414393ed6SStefano Zampini } 174514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 174614393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 174714393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 174814393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 174914393ed6SStefano Zampini } 175014393ed6SStefano Zampini 1751ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1752ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1753ffd830a3SStefano Zampini } else { 1754ca92afb2SStefano Zampini PetscScalar *av,*array; 1755ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1756ca92afb2SStefano Zampini PetscInt n; 1757ca92afb2SStefano Zampini PetscBool flg_row; 1758ffd830a3SStefano Zampini 1759ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1760ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1761ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1762ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1763ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1764ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1765ca92afb2SStefano Zampini PetscInt j; 1766ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1767ffd830a3SStefano Zampini } 1768ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1769ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1770ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1771ffd830a3SStefano Zampini } 1772ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 177314393ed6SStefano Zampini if (F) { 1774a3df083aSStefano Zampini if (need_benign_correction) { 1775df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1776a3df083aSStefano Zampini PetscScalar *marr; 1777a3df083aSStefano Zampini 1778a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 177914393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 178014393ed6SStefano Zampini 178114393ed6SStefano Zampini | 0 0 0 | (V) 178214393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 178314393ed6SStefano Zampini | 0 0 -1 | (p0) 178414393ed6SStefano Zampini 178514393ed6SStefano Zampini */ 1786df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 178714393ed6SStefano Zampini const PetscScalar *vals; 178814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 178914393ed6SStefano Zampini PetscInt n,j,nz; 179014393ed6SStefano Zampini 1791df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1792df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 179314393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 179414393ed6SStefano Zampini for (j=0;j<n;j++) { 179514393ed6SStefano Zampini PetscScalar val = vals[j]; 179614393ed6SStefano Zampini PetscInt k,col = idxs[j]; 179714393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 179814393ed6SStefano Zampini } 179914393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1800df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 180114393ed6SStefano Zampini } 180214393ed6SStefano Zampini /* need to correct the rhs */ 1803a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1804a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1805df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1806a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1807a3df083aSStefano Zampini } 1808a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1809a3df083aSStefano Zampini } 181006656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 181114393ed6SStefano Zampini /* need to correct the solution */ 1812a3df083aSStefano Zampini if (need_benign_correction) { 1813df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1814a3df083aSStefano Zampini PetscScalar *marr; 1815a3df083aSStefano Zampini 1816a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1817a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1818a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1819df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1820a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1821a3df083aSStefano Zampini } 1822a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1823a3df083aSStefano Zampini } 182406656605SStefano Zampini } else { 182506656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 182606656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1827ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1828ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 182906656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 183006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 183106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 183206656605SStefano Zampini } 183306656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 183406656605SStefano Zampini } 183580677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1836ffd830a3SStefano Zampini /* S_VV and S_CV */ 183706656605SStefano Zampini if (n_constraints) { 183806656605SStefano Zampini Mat B; 183980677318SStefano Zampini 1840ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 184180677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1842ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1843ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 184480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 184580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 184680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 184780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 184880677318SStefano Zampini } 1849ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 185080677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 185180677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1852ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 185380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 185406656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1855ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1856ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 185706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 185806656605SStefano Zampini } 185904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 186004708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 186104708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 186204708bb6SStefano Zampini } 1863ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1864ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1865ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1866ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1867ffd830a3SStefano Zampini } 186806656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 186914393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 187014393ed6SStefano Zampini if (need_benign_correction) { 1871df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 187214393ed6SStefano Zampini PetscScalar *marr,*sums; 187314393ed6SStefano Zampini 187414393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 187514393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1876df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 187714393ed6SStefano Zampini const PetscScalar *vals; 187814393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 187914393ed6SStefano Zampini PetscInt n,j,nz; 188014393ed6SStefano Zampini 1881df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1882df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 188314393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 188414393ed6SStefano Zampini PetscInt k; 188514393ed6SStefano Zampini sums[j] = 0.; 188614393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 188714393ed6SStefano Zampini } 188814393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 188914393ed6SStefano Zampini for (j=0;j<n;j++) { 189014393ed6SStefano Zampini PetscScalar val = vals[j]; 189114393ed6SStefano Zampini PetscInt k; 189214393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 189314393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 189414393ed6SStefano Zampini } 189514393ed6SStefano Zampini } 189614393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1897df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 189814393ed6SStefano Zampini } 189914393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 190014393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 190114393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 190214393ed6SStefano Zampini } 190380677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 190406656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 190506656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 190606656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 190706656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 190806656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 190906656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 191006656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1911d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1912019a44ceSStefano Zampini } else { 1913d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1914d16cbb6bSStefano Zampini } 19154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1916019a44ceSStefano Zampini const PetscScalar *vals; 1917019a44ceSStefano Zampini const PetscInt *idxs; 19184f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1919019a44ceSStefano Zampini 19204f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 19214f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1922d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 19234f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 19244f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1925019a44ceSStefano Zampini } 19264f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 192716f15bc4SStefano Zampini } 192821eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1929d16cbb6bSStefano Zampini 193006656605SStefano Zampini /* coarse basis functions */ 193106656605SStefano Zampini for (i=0;i<n_vertices;i++) { 193216f15bc4SStefano Zampini PetscScalar *y; 193316f15bc4SStefano Zampini 1934ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 193506656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 193606656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 193706656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193806656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193906656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 194006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 194106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 194206656605SStefano Zampini 194306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19444f1b2e48SStefano Zampini PetscInt j; 19454f1b2e48SStefano Zampini 194606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 194706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 194806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19514f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 195206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 195306656605SStefano Zampini } 195406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 195506656605SStefano Zampini } 195604708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 195704708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 195806656605SStefano Zampini } 195906656605SStefano Zampini 196006656605SStefano Zampini if (n_constraints) { 196106656605SStefano Zampini Mat B; 196206656605SStefano Zampini 1963ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 196406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 196580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 196606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 196706656605SStefano Zampini if (n_vertices) { 196880677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 196980677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 197080677318SStefano Zampini } else { 197180677318SStefano Zampini Mat S_VCt; 197280677318SStefano Zampini 1973ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1974ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1975ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1976ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1977ffd830a3SStefano Zampini } 197880677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 197980677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 198080677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 198180677318SStefano Zampini } 198206656605SStefano Zampini } 198306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 198406656605SStefano Zampini /* coarse basis functions */ 198506656605SStefano Zampini for (i=0;i<n_constraints;i++) { 198606656605SStefano Zampini PetscScalar *y; 198706656605SStefano Zampini 1988ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 198906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 199006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 199106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 199406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 199506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19964f1b2e48SStefano Zampini PetscInt j; 19974f1b2e48SStefano Zampini 199806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 199906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 200006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20034f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 200406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 200506656605SStefano Zampini } 200606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 200706656605SStefano Zampini } 200806656605SStefano Zampini } 200980677318SStefano Zampini if (n_constraints) { 201080677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 201180677318SStefano Zampini } 20124f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 2013019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 2014019a44ceSStefano Zampini 201506656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 20163301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2017ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2018ffd830a3SStefano Zampini PetscScalar *marray; 201906656605SStefano Zampini 202006656605SStefano Zampini if (n_constraints) { 2021ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 202206656605SStefano Zampini 2023ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 202406656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2025ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 202616f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 202706656605SStefano Zampini if (n_vertices) { 2028ffd830a3SStefano Zampini Mat S_VCT; 202906656605SStefano Zampini 203006656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2031ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 203216f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 203306656605SStefano Zampini } 2034ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 203506656605SStefano Zampini } 203616f15bc4SStefano Zampini if (n_vertices && n_R) { 2037ffd830a3SStefano Zampini PetscScalar *av,*marray; 2038ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2039ffd830a3SStefano Zampini PetscInt n; 2040ffd830a3SStefano Zampini PetscBool flg_row; 204106656605SStefano Zampini 2042ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2043ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2044ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2045ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2046ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2047ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2048ffd830a3SStefano Zampini PetscInt j; 2049ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2050ffd830a3SStefano Zampini } 2051ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2052ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2053ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 205406656605SStefano Zampini } 205506656605SStefano Zampini 2056ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2057ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2058ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2059ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2060ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 206106656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 206206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 206306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 206406656605SStefano Zampini } 2065ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2066ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2067ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2068ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2069ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2070ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2071ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2072ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 207306656605SStefano Zampini } 2074ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 207506656605SStefano Zampini /* coarse basis functions */ 207606656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 207706656605SStefano Zampini PetscScalar *y; 207806656605SStefano Zampini 2079ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 208006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 208106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 208206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208406656605SStefano Zampini if (i<n_vertices) { 208506656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 208606656605SStefano Zampini } 208706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 208806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 208906656605SStefano Zampini 209006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 209106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 209206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 209306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 209406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 209506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 209606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 209706656605SStefano Zampini } 209806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209906656605SStefano Zampini } 2100ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2101ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 210206656605SStefano Zampini } 2103d62866d3SStefano Zampini /* free memory */ 210488ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 210506656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 210606656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 210706656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 210806656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2109d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2110d62866d3SStefano Zampini if (n_vertices) { 2111d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2112d62866d3SStefano Zampini } 2113d62866d3SStefano Zampini if (n_constraints) { 2114d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2115d62866d3SStefano Zampini } 211688ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 211788ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 211888ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2119d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 212088ebb749SStefano Zampini Mat coarse_sub_mat; 212125084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 212288ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 212388ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 212488ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 21258bec7fa6SStefano Zampini Mat C_B,CPHI; 21268bec7fa6SStefano Zampini IS is_dummy; 21278bec7fa6SStefano Zampini Vec mones; 212888ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 212988ebb749SStefano Zampini PetscReal real_value; 213088ebb749SStefano Zampini 2131a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2132a3df083aSStefano Zampini Mat A; 2133a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2134a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2135a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2136a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2137a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2138a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2139a3df083aSStefano Zampini } else { 214088ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 214188ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 214288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 214388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2144a3df083aSStefano Zampini } 214588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 214688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2147ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 214888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 214988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 215088ebb749SStefano Zampini } 215188ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 215288ebb749SStefano Zampini 215325084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 21543301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 215525084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2156ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 215788ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 215888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 215988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216088ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 216188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 216288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 216488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 216588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 216788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 216888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216988ebb749SStefano Zampini } else { 217088ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 217188ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 217288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 217388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 217488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 217588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 217688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 217788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 217888ebb749SStefano Zampini } 217988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 218088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 218188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 218288ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 21834f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2184d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 2185d12edf2fSStefano Zampini PetscScalar *data,*data2; 21864f1b2e48SStefano Zampini PetscInt j; 2187d12edf2fSStefano Zampini 21884f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 21894f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 21904f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 2191d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2192d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2193d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2194d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 21954f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 21964f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2197d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 21984f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 21994f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22004f1b2e48SStefano Zampini } 2201d12edf2fSStefano Zampini } 2202d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2203d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2204d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2205d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2206d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2207d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 2208d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 2209d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2210d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 2211d12edf2fSStefano Zampini } 2212d12edf2fSStefano Zampini #if 0 2213d12edf2fSStefano Zampini { 2214d12edf2fSStefano Zampini PetscViewer viewer; 2215d12edf2fSStefano Zampini char filename[256]; 2216ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2217d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2218d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2219ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2220ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2221ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2222d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 2223ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2224ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 2225ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2226ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2227ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2228ffd830a3SStefano Zampini } 2229ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2230ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2231ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2232ffd830a3SStefano Zampini } 2233ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 2234ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2235ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2236ffd830a3SStefano Zampini } 2237d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2238d12edf2fSStefano Zampini } 2239d12edf2fSStefano Zampini #endif 224081d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 22418bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 22421575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 224306656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 22448bec7fa6SStefano Zampini 22458bec7fa6SStefano Zampini /* check constraints */ 22464f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 22478bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 22488bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 22498bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 22508bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 22518bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 22528bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 22538bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2254bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2255ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2256bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2257bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2258bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2259bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2260bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 226188ebb749SStefano Zampini } 22628bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 22638bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 22648bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 22658bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2266d12edf2fSStefano Zampini } 226725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 226888ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 226988ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 227088ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 227188ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 227288ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 227688ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 227788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2278ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 227988ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 228088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 228188ebb749SStefano Zampini } 228288ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 228388ebb749SStefano Zampini } 22848629588bSStefano Zampini /* get back data */ 22858629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 228688ebb749SStefano Zampini PetscFunctionReturn(0); 228788ebb749SStefano Zampini } 228888ebb749SStefano Zampini 228988ebb749SStefano Zampini #undef __FUNCT__ 2290d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2291d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2292aa0d41d4SStefano Zampini { 2293d65f70fdSStefano Zampini Mat *work_mat; 2294d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2295d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2296d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2297aa0d41d4SStefano Zampini PetscErrorCode ierr; 2298aa0d41d4SStefano Zampini 2299aa0d41d4SStefano Zampini PetscFunctionBegin; 2300d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2301d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2302d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2303d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2304aa0d41d4SStefano Zampini 2305d65f70fdSStefano Zampini if (!rsorted) { 2306906d46d4SStefano Zampini const PetscInt *idxs; 2307906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2308aa0d41d4SStefano Zampini 2309d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2310d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2311d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2312d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2313aa0d41d4SStefano Zampini } 2314d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2315d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2316d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2317d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2318aa0d41d4SStefano Zampini } 2319d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2320d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2321d65f70fdSStefano Zampini } else { 2322d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2323d65f70fdSStefano Zampini isrow_s = isrow; 2324aa0d41d4SStefano Zampini } 2325906d46d4SStefano Zampini 2326d65f70fdSStefano Zampini if (!csorted) { 2327d65f70fdSStefano Zampini if (isrow == iscol) { 2328d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2329d65f70fdSStefano Zampini iscol_s = isrow_s; 2330d65f70fdSStefano Zampini } else { 2331d65f70fdSStefano Zampini const PetscInt *idxs; 2332d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2333906d46d4SStefano Zampini 2334d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2335d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2336d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2337d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2338d65f70fdSStefano Zampini } 2339d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2340d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2341d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2342d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2343d65f70fdSStefano Zampini } 2344d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2345d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2346d65f70fdSStefano Zampini } 2347d65f70fdSStefano Zampini } else { 2348d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2349d65f70fdSStefano Zampini iscol_s = iscol; 2350d65f70fdSStefano Zampini } 2351d65f70fdSStefano Zampini 2352d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2353d65f70fdSStefano Zampini 2354d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2355906d46d4SStefano Zampini Mat new_mat; 2356d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2357906d46d4SStefano Zampini 2358d65f70fdSStefano Zampini if (!rsorted) { 2359d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2360d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2361d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2362d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2363906d46d4SStefano Zampini } 2364d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2365d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2366d65f70fdSStefano Zampini } else { 2367d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2368906d46d4SStefano Zampini } 2369d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2370d65f70fdSStefano Zampini 2371d65f70fdSStefano Zampini if (!csorted) { 2372d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2373d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2374d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2375d65f70fdSStefano Zampini } else { 2376d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2377d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2378d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2379d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2380d65f70fdSStefano Zampini } 2381d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2382d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2383d65f70fdSStefano Zampini } 2384d65f70fdSStefano Zampini } else { 2385d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2386d65f70fdSStefano Zampini } 2387d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2388d65f70fdSStefano Zampini 2389d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2390d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2391d65f70fdSStefano Zampini work_mat[0] = new_mat; 2392d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2393d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2394d65f70fdSStefano Zampini } 2395d65f70fdSStefano Zampini 2396d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2397d65f70fdSStefano Zampini *B = work_mat[0]; 2398d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2399d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2400d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2401d65f70fdSStefano Zampini PetscFunctionReturn(0); 2402d65f70fdSStefano Zampini } 2403d65f70fdSStefano Zampini 2404d65f70fdSStefano Zampini #undef __FUNCT__ 24055e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 24065e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2407aa0d41d4SStefano Zampini { 2408aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 24095e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2410d65f70fdSStefano Zampini Mat new_mat; 24115e8657edSStefano Zampini IS is_local,is_global; 2412d65f70fdSStefano Zampini PetscInt local_size; 2413d65f70fdSStefano Zampini PetscBool isseqaij; 2414aa0d41d4SStefano Zampini PetscErrorCode ierr; 2415aa0d41d4SStefano Zampini 2416aa0d41d4SStefano Zampini PetscFunctionBegin; 2417aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 24185e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 24195e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2420b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2421aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2422d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2423aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2424906d46d4SStefano Zampini 2425906d46d4SStefano Zampini /* check */ 2426906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2427906d46d4SStefano Zampini Vec x,x_change; 2428906d46d4SStefano Zampini PetscReal error; 2429906d46d4SStefano Zampini 24305e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2431906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 24325e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2433e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2434e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2435d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2436e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2437e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2438906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2439906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2440906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2441906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2442906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2443906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2444906d46d4SStefano Zampini } 2445906d46d4SStefano Zampini 244622d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 24479b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 244822d5777bSStefano Zampini if (isseqaij) { 24491cf9b237SStefano Zampini Mat M; 24501cf9b237SStefano Zampini 24511cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24521cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24531cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2454aa0d41d4SStefano Zampini } else { 24551cf9b237SStefano Zampini Mat work_mat,M; 24561cf9b237SStefano Zampini 2457aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 24581cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24591cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24601cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2461aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2462aa0d41d4SStefano Zampini } 24633301b35fSStefano Zampini if (matis->A->symmetric_set) { 24643301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2465e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24663301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2467e496cd5dSStefano Zampini #endif 24683301b35fSStefano Zampini } 246945a1bb75SStefano Zampini /* 247045a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2471d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 247245a1bb75SStefano Zampini */ 2473d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2474aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2475aa0d41d4SStefano Zampini } 2476aa0d41d4SStefano Zampini 2477aa0d41d4SStefano Zampini #undef __FUNCT__ 2478a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 24798ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2480a64d13efSStefano Zampini { 2481a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2482a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2483d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 248453892102SStefano Zampini PetscInt *idx_R_local=NULL; 24853a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 24863a50541eSStefano Zampini PetscInt vbs,bs; 24876816873aSStefano Zampini PetscBT bitmask=NULL; 2488a64d13efSStefano Zampini PetscErrorCode ierr; 2489a64d13efSStefano Zampini 2490a64d13efSStefano Zampini PetscFunctionBegin; 2491b23d619eSStefano Zampini /* 2492b23d619eSStefano Zampini No need to setup local scatters if 2493b23d619eSStefano Zampini - primal space is unchanged 2494b23d619eSStefano Zampini AND 2495b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2496b23d619eSStefano Zampini AND 2497b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2498b23d619eSStefano Zampini */ 2499b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2500f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2501f4ddd8eeSStefano Zampini } 2502f4ddd8eeSStefano Zampini /* destroy old objects */ 2503f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2504f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2505f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2506a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2507b371cd4fSStefano Zampini n_B = pcis->n_B; 2508b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2509b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 25103a50541eSStefano Zampini 2511a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 25126816873aSStefano Zampini 251353892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2514df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2515854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2516a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2517a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25180e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2519a64d13efSStefano Zampini } 2520a64d13efSStefano Zampini 2521a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 25224641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 25236816873aSStefano Zampini idx_R_local[n_R++] = i; 2524a64d13efSStefano Zampini } 2525a64d13efSStefano Zampini } 2526df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2527df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 25286816873aSStefano Zampini 2529df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2530df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 25316816873aSStefano Zampini } 25323a50541eSStefano Zampini 25333a50541eSStefano Zampini /* Block code */ 25343a50541eSStefano Zampini vbs = 1; 25353a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 25363a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 25373a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 25383a50541eSStefano Zampini PetscInt *vary; 2539df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2540785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 25413a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2542d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2543d3df7717SStefano 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 */ 25440e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2545d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 25463a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 25473a50541eSStefano Zampini is_blocked = PETSC_FALSE; 25483a50541eSStefano Zampini break; 25493a50541eSStefano Zampini } 25503a50541eSStefano Zampini } 2551d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2552d3df7717SStefano Zampini } else { 2553d3df7717SStefano Zampini /* Verify directly the R set */ 2554d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2555d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2556d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2557d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2558d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2559d3df7717SStefano Zampini break; 2560d3df7717SStefano Zampini } 2561d3df7717SStefano Zampini } 2562d3df7717SStefano Zampini } 2563d3df7717SStefano Zampini } 25643a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 25653a50541eSStefano Zampini vbs = bs; 25663a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 25673a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 25683a50541eSStefano Zampini } 25693a50541eSStefano Zampini } 25703a50541eSStefano Zampini } 25713a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2572df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2573df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 257453892102SStefano Zampini 2575df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2576df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 257753892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2578df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 257953892102SStefano Zampini } else { 25803a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 258153892102SStefano Zampini } 2582a64d13efSStefano Zampini 2583a64d13efSStefano Zampini /* print some info if requested */ 2584a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2585a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2586a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 25871575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2588a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2589a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 25904f1b2e48SStefano 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); 2591a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2592a64d13efSStefano Zampini } 2593a64d13efSStefano Zampini 2594a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2595df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 25966816873aSStefano Zampini IS is_aux1,is_aux2; 25976816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 25986816873aSStefano Zampini 25993a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2600854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2601854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2602a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26034641a718SStefano Zampini for (i=0; i<n_D; i++) { 26044641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26054641a718SStefano Zampini } 2606a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2607a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26084641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 26094641a718SStefano Zampini aux_array1[j++] = i; 2610a64d13efSStefano Zampini } 2611a64d13efSStefano Zampini } 2612a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2613a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2614a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 26154641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 26164641a718SStefano Zampini aux_array2[j++] = i; 2617a64d13efSStefano Zampini } 2618a64d13efSStefano Zampini } 2619a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2620a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2621a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2622a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2623a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2624a64d13efSStefano Zampini 26258eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2626785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2627a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26284641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 26294641a718SStefano Zampini aux_array1[j++] = i; 2630a64d13efSStefano Zampini } 2631a64d13efSStefano Zampini } 2632a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2633a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2634a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2635a64d13efSStefano Zampini } 26364641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 26373a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2638d62866d3SStefano Zampini } else { 2639df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26406816873aSStefano Zampini IS tis; 26416816873aSStefano Zampini PetscInt schur_size; 26426816873aSStefano Zampini 2643df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 26446816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2645df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 26466816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 26476816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 26486816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 26496816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 26506816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2651d62866d3SStefano Zampini } 2652d62866d3SStefano Zampini } 2653a64d13efSStefano Zampini PetscFunctionReturn(0); 2654a64d13efSStefano Zampini } 2655a64d13efSStefano Zampini 2656304d26faSStefano Zampini 2657304d26faSStefano Zampini #undef __FUNCT__ 2658304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2659684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2660304d26faSStefano Zampini { 2661304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2662304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2663304d26faSStefano Zampini PC pc_temp; 2664304d26faSStefano Zampini Mat A_RR; 2665f4ddd8eeSStefano Zampini MatReuse reuse; 2666304d26faSStefano Zampini PetscScalar m_one = -1.0; 2667304d26faSStefano Zampini PetscReal value; 266804708bb6SStefano Zampini PetscInt n_D,n_R; 26699577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2670304d26faSStefano Zampini PetscErrorCode ierr; 2671e604994aSStefano Zampini /* prefixes stuff */ 2672312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2673e604994aSStefano Zampini size_t len; 2674304d26faSStefano Zampini 2675304d26faSStefano Zampini PetscFunctionBegin; 2676304d26faSStefano Zampini 2677e604994aSStefano Zampini /* compute prefixes */ 2678e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2679e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2680e604994aSStefano Zampini if (!pcbddc->current_level) { 2681e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2682e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2683e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2684e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2685e604994aSStefano Zampini } else { 2686e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2687312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2688e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2689e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2690312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2691312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 269234d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 269334d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2694e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2695e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2696e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2697e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2698e604994aSStefano Zampini } 2699e604994aSStefano Zampini 2700304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2701684f6988SStefano Zampini if (dirichlet) { 2702d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2703a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2704a3df083aSStefano Zampini Mat A_IIn; 2705a3df083aSStefano Zampini 2706a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2707a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2708a3df083aSStefano Zampini pcis->A_II = A_IIn; 2709a3df083aSStefano Zampini } 27103301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27113301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2712964fefecSStefano Zampini } 2713ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2714964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2715304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2716304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2717304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2718304d26faSStefano Zampini /* default */ 2719304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2720e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 27219577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2722304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 27239577ea80SStefano Zampini if (issbaij) { 27249577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 27259577ea80SStefano Zampini } else { 2726304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 27279577ea80SStefano Zampini } 2728304d26faSStefano Zampini /* Allow user's customization */ 2729304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2730304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2731304d26faSStefano Zampini } 2732d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2733df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2734df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2735d62866d3SStefano Zampini 2736df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2737d5574798SStefano Zampini } 2738304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2739304d26faSStefano Zampini if (!n_D) { 2740304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2741304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2742304d26faSStefano Zampini } 2743304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2744304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2745304d26faSStefano Zampini /* set ksp_D into pcis data */ 2746304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2747304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2748304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2749684f6988SStefano Zampini } 2750304d26faSStefano Zampini 2751304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2752684f6988SStefano Zampini A_RR = 0; 2753684f6988SStefano Zampini if (neumann) { 2754d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 275504708bb6SStefano Zampini PetscInt ibs,mbs; 275604708bb6SStefano Zampini PetscBool issbaij; 275704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2758f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 27598ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2760f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2761f4ddd8eeSStefano Zampini PetscInt nn_R; 276281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2763f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2764f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2765f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2766f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2767f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2768f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2769f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2770727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2771f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2772f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2773f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2774f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2775f4ddd8eeSStefano Zampini } 2776f4ddd8eeSStefano Zampini } 2777f4ddd8eeSStefano Zampini /* last check */ 2778d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2779f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2780f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2781f4ddd8eeSStefano Zampini } 2782f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2783f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2784f4ddd8eeSStefano Zampini } 2785f4ddd8eeSStefano Zampini /* extract A_RR */ 2786af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2787af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 278804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 278904708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 279004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 279104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 279204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2793af732b37SStefano Zampini } else { 279404708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 27956816873aSStefano Zampini } 279604708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 279704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 279804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 279904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 280004708bb6SStefano Zampini } else { 280104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 280204708bb6SStefano Zampini } 280304708bb6SStefano Zampini } 2804f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 28053301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28063301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 28076816873aSStefano Zampini } 2808a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2809a3df083aSStefano Zampini Mat A_RRn; 2810a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RRn);CHKERRQ(ierr); 2811a3df083aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2812a3df083aSStefano Zampini A_RR = A_RRn; 2813a3df083aSStefano Zampini } 2814f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2815304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2816304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2817304d26faSStefano Zampini /* default */ 2818304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2819e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2820304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 28219577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 28229577ea80SStefano Zampini if (issbaij) { 28239577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28249577ea80SStefano Zampini } else { 2825304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28269577ea80SStefano Zampini } 2827304d26faSStefano Zampini /* Allow user's customization */ 2828304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2829304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2830304d26faSStefano Zampini } 2831304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2832304d26faSStefano Zampini if (!n_R) { 2833304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2834304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2835304d26faSStefano Zampini } 2836df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2837df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2838df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2839d62866d3SStefano Zampini 2840df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2841d62866d3SStefano Zampini } 2842ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2843304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2844304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2845684f6988SStefano Zampini } 28466816873aSStefano Zampini /* free Neumann problem's matrix */ 28476816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2848304d26faSStefano Zampini 2849304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 28500fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2851684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2852684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28531575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2854684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2855684f6988SStefano Zampini } 2856684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 28570fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 28580fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 28590fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 28600fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 28610fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2862304d26faSStefano Zampini /* need to be adapted? */ 2863b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2864b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2865b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2866304d26faSStefano Zampini /* print info */ 2867304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2868e604994aSStefano 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); 2869304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2870304d26faSStefano Zampini } 2871b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2872298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2873304d26faSStefano Zampini } 2874684f6988SStefano Zampini } 2875684f6988SStefano Zampini if (neumann) { /* Neumann */ 28766816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 28770fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 28780fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 28790fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 28800fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 28810fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2882304d26faSStefano Zampini /* need to be adapted? */ 2883b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2884b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2885304d26faSStefano Zampini /* print info */ 2886304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2887e604994aSStefano 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); 2888304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2889304d26faSStefano Zampini } 2890b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2891298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2892304d26faSStefano Zampini } 28930fccc4e9SStefano Zampini } 2894684f6988SStefano Zampini } 2895304d26faSStefano Zampini PetscFunctionReturn(0); 2896304d26faSStefano Zampini } 2897304d26faSStefano Zampini 2898304d26faSStefano Zampini #undef __FUNCT__ 2899ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 290080677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2901674ae819SStefano Zampini { 2902674ae819SStefano Zampini PetscErrorCode ierr; 2903674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2904be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2905674ae819SStefano Zampini 2906674ae819SStefano Zampini PetscFunctionBegin; 2907df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 290880677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 290920c7b377SStefano Zampini } 291080677318SStefano Zampini if (!pcbddc->switch_static) { 291180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 291280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 291380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 291420c7b377SStefano Zampini } 2915df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 291680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 291780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 291820c7b377SStefano Zampini } else { 2919df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2920be83ff47SStefano Zampini 2921df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2922df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 292320c7b377SStefano Zampini } 2924be83ff47SStefano Zampini } else { 292580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 292680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 292780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 292880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 292980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 293080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 293180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 293280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 293380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2934674ae819SStefano Zampini } 2935674ae819SStefano Zampini } 2936df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 293780677318SStefano Zampini if (applytranspose) { 293880677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 293980677318SStefano Zampini } else { 294080677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 294180677318SStefano Zampini } 2942be83ff47SStefano Zampini } else { 2943df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2944be83ff47SStefano Zampini 2945be83ff47SStefano Zampini if (applytranspose) { 2946df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2947be83ff47SStefano Zampini } else { 2948df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2949be83ff47SStefano Zampini } 2950be83ff47SStefano Zampini } 295180677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 295280677318SStefano Zampini if (!pcbddc->switch_static) { 2953df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 295480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2956be83ff47SStefano Zampini } else { 2957df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2958be83ff47SStefano Zampini 2959df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2960df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2961be83ff47SStefano Zampini } 296280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 296380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 296480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 296580677318SStefano Zampini } 296680677318SStefano Zampini } else { 296780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 296880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 296980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 297280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 297380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 297480677318SStefano Zampini } 297580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2979674ae819SStefano Zampini } 2980674ae819SStefano Zampini PetscFunctionReturn(0); 2981674ae819SStefano Zampini } 2982674ae819SStefano Zampini 2983dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2984674ae819SStefano Zampini #undef __FUNCT__ 2985674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2986dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2987674ae819SStefano Zampini { 2988674ae819SStefano Zampini PetscErrorCode ierr; 2989674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2990674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2991674ae819SStefano Zampini const PetscScalar zero = 0.0; 2992674ae819SStefano Zampini 2993674ae819SStefano Zampini PetscFunctionBegin; 2994dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2995dc359a40SStefano Zampini if (applytranspose) { 2996674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 29978eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2998dc359a40SStefano Zampini } else { 2999674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3000674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 300115aaf578SStefano Zampini } 3002efc2fbd9SStefano Zampini 3003efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 30044f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3005efc2fbd9SStefano Zampini PetscScalar *array; 30064f1b2e48SStefano Zampini PetscInt j; 3007efc2fbd9SStefano Zampini 3008efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30094f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3010efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3011efc2fbd9SStefano Zampini } 3012efc2fbd9SStefano Zampini 301312edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 301412edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 301512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 301612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 301712edc857SStefano Zampini 30189f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 301912edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 302012edc857SStefano Zampini if (pcbddc->coarse_ksp) { 302151694757SStefano Zampini Mat coarse_mat; 3022964fefecSStefano Zampini Vec rhs,sol; 302351694757SStefano Zampini MatNullSpace nullsp; 3024964fefecSStefano Zampini 3025964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3026964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 302751694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 302851694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 302951694757SStefano Zampini if (nullsp) { 303051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 303151694757SStefano Zampini } 303212edc857SStefano Zampini if (applytranspose) { 3033964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 303412edc857SStefano Zampini } else { 3035964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 303612edc857SStefano Zampini } 303751694757SStefano Zampini if (nullsp) { 303851694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 303951694757SStefano Zampini } 304012edc857SStefano Zampini } 3041674ae819SStefano Zampini 3042674ae819SStefano Zampini /* Local solution on R nodes */ 304380677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 304480677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 30459f00e9b4SStefano Zampini } 3046674ae819SStefano Zampini 30479f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 30489f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3050674ae819SStefano Zampini 3051674ae819SStefano Zampini /* Sum contributions from two levels */ 3052dc359a40SStefano Zampini if (applytranspose) { 3053dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3054dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3055dc359a40SStefano Zampini } else { 3056674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 30578eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3058dc359a40SStefano Zampini } 3059efc2fbd9SStefano Zampini /* store p0 */ 30604f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3061efc2fbd9SStefano Zampini PetscScalar *array; 30624f1b2e48SStefano Zampini PetscInt j; 3063efc2fbd9SStefano Zampini 3064efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30654f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3066efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3067efc2fbd9SStefano Zampini } 3068674ae819SStefano Zampini PetscFunctionReturn(0); 3069674ae819SStefano Zampini } 3070674ae819SStefano Zampini 3071674ae819SStefano Zampini #undef __FUNCT__ 3072674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 307312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3074674ae819SStefano Zampini { 3075674ae819SStefano Zampini PetscErrorCode ierr; 3076674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 307758da7f69SStefano Zampini PetscScalar *array; 307812edc857SStefano Zampini Vec from,to; 3079674ae819SStefano Zampini 3080674ae819SStefano Zampini PetscFunctionBegin; 308112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 308212edc857SStefano Zampini from = pcbddc->coarse_vec; 308312edc857SStefano Zampini to = pcbddc->vec1_P; 308412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 308512edc857SStefano Zampini Vec tvec; 308658da7f69SStefano Zampini 308758da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 308858da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 308912edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 309058da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 309158da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 309258da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 309312edc857SStefano Zampini } 309412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 309512edc857SStefano Zampini from = pcbddc->vec1_P; 309612edc857SStefano Zampini to = pcbddc->coarse_vec; 309712edc857SStefano Zampini } 309812edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3099674ae819SStefano Zampini PetscFunctionReturn(0); 3100674ae819SStefano Zampini } 3101674ae819SStefano Zampini 3102674ae819SStefano Zampini #undef __FUNCT__ 3103674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 310412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3105674ae819SStefano Zampini { 3106674ae819SStefano Zampini PetscErrorCode ierr; 3107674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 310858da7f69SStefano Zampini PetscScalar *array; 310912edc857SStefano Zampini Vec from,to; 3110674ae819SStefano Zampini 3111674ae819SStefano Zampini PetscFunctionBegin; 311212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 311312edc857SStefano Zampini from = pcbddc->coarse_vec; 311412edc857SStefano Zampini to = pcbddc->vec1_P; 311512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 311612edc857SStefano Zampini from = pcbddc->vec1_P; 311712edc857SStefano Zampini to = pcbddc->coarse_vec; 311812edc857SStefano Zampini } 311912edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 312012edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 312112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 312212edc857SStefano Zampini Vec tvec; 312358da7f69SStefano Zampini 312412edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 312558da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 312658da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 312758da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 312858da7f69SStefano Zampini } 312958da7f69SStefano Zampini } else { 313058da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 313158da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 313212edc857SStefano Zampini } 313312edc857SStefano Zampini } 3134674ae819SStefano Zampini PetscFunctionReturn(0); 3135674ae819SStefano Zampini } 3136674ae819SStefano Zampini 3137984c4197SStefano Zampini /* uncomment for testing purposes */ 3138984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3139674ae819SStefano Zampini #undef __FUNCT__ 3140674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3141674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3142674ae819SStefano Zampini { 3143674ae819SStefano Zampini PetscErrorCode ierr; 3144674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3145674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3146674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3147984c4197SStefano Zampini /* one and zero */ 3148984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3149984c4197SStefano Zampini /* space to store constraints and their local indices */ 31509162d606SStefano Zampini PetscScalar *constraints_data; 31519162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 31529162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 31539162d606SStefano Zampini PetscInt *constraints_n; 3154984c4197SStefano Zampini /* iterators */ 3155b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3156984c4197SStefano Zampini /* BLAS integers */ 3157e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3158e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3159c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3160727cdba6SStefano Zampini /* reuse */ 31610e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 31620e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3163984c4197SStefano Zampini /* change of basis */ 3164b3d85658SStefano Zampini PetscBool qr_needed; 31659162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3166984c4197SStefano Zampini /* auxiliary stuff */ 316764efe560SStefano Zampini PetscInt *nnz,*is_indices; 31688a0068c3SStefano Zampini PetscInt ncc; 3169984c4197SStefano Zampini /* some quantities */ 317045a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3171a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3172984c4197SStefano Zampini 3173674ae819SStefano Zampini PetscFunctionBegin; 31748e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 31758e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 31768e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3177088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3178088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 31790e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 31800e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 31810e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 31820e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 31830e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3184088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3185cf5a6209SStefano Zampini 3186cf5a6209SStefano Zampini /* print some info */ 3187cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3188cf5a6209SStefano Zampini IS vertices; 3189cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3190cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3191cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3192cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 31931575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3194cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3195cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3196fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3197fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3198cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31991575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3200cf5a6209SStefano Zampini } 3201cf5a6209SStefano Zampini 3202cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 32039162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3204cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3205cf5a6209SStefano Zampini const Vec *nearnullvecs; 3206cf5a6209SStefano Zampini Vec *localnearnullsp; 3207cf5a6209SStefano Zampini PetscScalar *array; 3208cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3209cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3210674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3211b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3212674ae819SStefano Zampini PetscScalar *work; 3213674ae819SStefano Zampini PetscReal *singular_vals; 3214674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3215674ae819SStefano Zampini PetscReal *rwork; 3216674ae819SStefano Zampini #endif 3217674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3218674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3219674ae819SStefano Zampini #else 3220964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3221964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3222674ae819SStefano Zampini #endif 3223674ae819SStefano Zampini 3224674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3225d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3226d06fc5fdSStefano Zampini /* free unneeded index sets */ 3227d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3228d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3229674ae819SStefano Zampini } 3230d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3231d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3232d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3233d06fc5fdSStefano Zampini } 3234d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3235d06fc5fdSStefano Zampini n_ISForEdges = 0; 3236d06fc5fdSStefano Zampini } 3237d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3238d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3239d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3240d06fc5fdSStefano Zampini } 3241d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3242d06fc5fdSStefano Zampini n_ISForFaces = 0; 3243d06fc5fdSStefano Zampini } 324470022509SStefano Zampini 324570022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 324670022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 324770022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 324870022509SStefano Zampini if (pcbddc->NullSpace) { 324970022509SStefano Zampini PetscBool tbool[2],gbool[2]; 325070022509SStefano Zampini 325170022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3252b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3253d06fc5fdSStefano Zampini if (!ISForEdges) { 3254d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3255d06fc5fdSStefano Zampini } 3256b8ffe317SStefano Zampini } 3257d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3258d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3259d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3260d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3261d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 326298a51de6SStefano Zampini } 326370022509SStefano Zampini #endif 326408122e43SStefano Zampini 3265674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3266674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3267674ae819SStefano Zampini if (nearnullsp) { 3268674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3269f4ddd8eeSStefano Zampini /* remove any stored info */ 3270f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3271f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3272f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3273f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3274f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3275473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3276f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3277f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3278f4ddd8eeSStefano Zampini } 3279984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3280984c4197SStefano Zampini nnsp_size = 0; 3281674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3282674ae819SStefano Zampini } 3283984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3284984c4197SStefano Zampini max_constraints = nnsp_size; 3285984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3286984c4197SStefano Zampini 3287674ae819SStefano Zampini /* 3288674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 32899162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 32909162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 32919162d606SStefano Zampini There can be multiple constraints per connected component 3292674ae819SStefano Zampini */ 3293674ae819SStefano Zampini n_vertices = 0; 3294674ae819SStefano Zampini if (ISForVertices) { 3295674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3296674ae819SStefano Zampini } 32979162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 32989162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 32999162d606SStefano Zampini 33009162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 33019162d606SStefano Zampini total_counts *= max_constraints; 3302674ae819SStefano Zampini total_counts += n_vertices; 33034641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 33049162d606SStefano Zampini 3305674ae819SStefano Zampini total_counts = 0; 3306674ae819SStefano Zampini max_size_of_constraint = 0; 3307674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 33089162d606SStefano Zampini IS used_is; 3309674ae819SStefano Zampini if (i<n_ISForEdges) { 33109162d606SStefano Zampini used_is = ISForEdges[i]; 3311674ae819SStefano Zampini } else { 33129162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3313674ae819SStefano Zampini } 33149162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3315674ae819SStefano Zampini total_counts += j; 3316674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3317674ae819SStefano Zampini } 33189162d606SStefano 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); 33199162d606SStefano Zampini 3320984c4197SStefano Zampini /* get local part of global near null space vectors */ 3321785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3322984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3323984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3324e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3325e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3326984c4197SStefano Zampini } 3327674ae819SStefano Zampini 3328242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3329242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3330a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3331242a89d7SStefano Zampini 3332984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3333a773dcb8SStefano Zampini if (!skip_lapack) { 3334674ae819SStefano Zampini PetscScalar temp_work; 3335911cabfeSStefano Zampini 3336674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3337984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3338785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3339785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3340785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3341674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3342785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3343674ae819SStefano Zampini #endif 3344674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3345c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3346c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3347674ae819SStefano Zampini lwork = -1; 3348674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3349674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3350c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3351674ae819SStefano Zampini #else 3352c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3353674ae819SStefano Zampini #endif 3354674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3355984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3356674ae819SStefano Zampini #else /* on missing GESVD */ 3357674ae819SStefano Zampini /* SVD */ 3358674ae819SStefano Zampini PetscInt max_n,min_n; 3359674ae819SStefano Zampini max_n = max_size_of_constraint; 3360984c4197SStefano Zampini min_n = max_constraints; 3361984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3362674ae819SStefano Zampini min_n = max_size_of_constraint; 3363984c4197SStefano Zampini max_n = max_constraints; 3364674ae819SStefano Zampini } 3365785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3366674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3367785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3368674ae819SStefano Zampini #endif 3369674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3370674ae819SStefano Zampini lwork = -1; 3371e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3372e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3373b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3374674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3375674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 33769162d606SStefano 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)); 3377674ae819SStefano Zampini #else 33789162d606SStefano 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)); 3379674ae819SStefano Zampini #endif 3380674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3381984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3382984c4197SStefano Zampini #endif /* on missing GESVD */ 3383674ae819SStefano Zampini /* Allocate optimal workspace */ 3384674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3385854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3386674ae819SStefano Zampini } 3387674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3388674ae819SStefano Zampini total_counts = 0; 33899162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 33909162d606SStefano Zampini constraints_data_ptr[0] = 0; 3391674ae819SStefano Zampini /* vertices */ 33929162d606SStefano Zampini if (n_vertices) { 3393674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 33949162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3395674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 33969162d606SStefano Zampini constraints_n[total_counts] = 1; 33979162d606SStefano Zampini constraints_data[total_counts] = 1.0; 33989162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 33999162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3400674ae819SStefano Zampini total_counts++; 3401674ae819SStefano Zampini } 3402674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3403674ae819SStefano Zampini n_vertices = total_counts; 3404674ae819SStefano Zampini } 3405984c4197SStefano Zampini 3406674ae819SStefano Zampini /* edges and faces */ 34079162d606SStefano Zampini total_counts_cc = total_counts; 3408911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 34099162d606SStefano Zampini IS used_is; 34109162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 34119162d606SStefano Zampini 3412911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 34139162d606SStefano Zampini used_is = ISForEdges[ncc]; 3414984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3415674ae819SStefano Zampini } else { 34169162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3417984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3418674ae819SStefano Zampini } 3419674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 34209162d606SStefano Zampini 34219162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 34229162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3423984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3424984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3425674ae819SStefano Zampini if (nnsp_has_cnst) { 34265b08dc53SStefano Zampini PetscScalar quad_value; 34279162d606SStefano Zampini 34289162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34299162d606SStefano Zampini idxs_copied = PETSC_TRUE; 34309162d606SStefano Zampini 3431a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3432674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3433a773dcb8SStefano Zampini } else { 3434a773dcb8SStefano Zampini quad_value = 1.0; 3435a773dcb8SStefano Zampini } 3436674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34379162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3438674ae819SStefano Zampini } 34399162d606SStefano Zampini temp_constraints++; 3440674ae819SStefano Zampini total_counts++; 3441674ae819SStefano Zampini } 3442674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3443984c4197SStefano Zampini PetscReal real_value; 34449162d606SStefano Zampini PetscScalar *ptr_to_data; 34459162d606SStefano Zampini 3446984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 34479162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3448674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34499162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3450674ae819SStefano Zampini } 3451984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3452984c4197SStefano Zampini /* check if array is null on the connected component */ 3453e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34549162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 34555b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3456674ae819SStefano Zampini temp_constraints++; 3457674ae819SStefano Zampini total_counts++; 34589162d606SStefano Zampini if (!idxs_copied) { 34599162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34609162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3461674ae819SStefano Zampini } 3462674ae819SStefano Zampini } 34639162d606SStefano Zampini } 34649162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 346545a1bb75SStefano Zampini valid_constraints = temp_constraints; 3466eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3467a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 34689162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 34699162d606SStefano Zampini 34709162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3471a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34729162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3473a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 34749162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3475a773dcb8SStefano Zampini } else { /* perform SVD */ 3476984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 34779162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3478674ae819SStefano Zampini 3479674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3480984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3481984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3482984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3483984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3484984c4197SStefano Zampini from that computed using LAPACKgesvd 3485984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3486984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3487984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3488674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3489e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3490984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3491674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3492674ae819SStefano Zampini for (k=0;k<j+1;k++) { 34939162d606SStefano 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)); 3494674ae819SStefano Zampini } 3495674ae819SStefano Zampini } 3496e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3497e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3498e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3499674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3500c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3501674ae819SStefano Zampini #else 3502c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3503674ae819SStefano Zampini #endif 3504674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3505984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3506984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3507674ae819SStefano Zampini j = 0; 3508984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3509674ae819SStefano Zampini total_counts = total_counts-j; 351045a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3511e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3512c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3513c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3514c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3515c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3516c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3517c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3518674ae819SStefano Zampini if (j<temp_constraints) { 3519984c4197SStefano Zampini PetscInt ii; 3520984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3521674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 35229162d606SStefano 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)); 3523674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3524984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3525674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 35269162d606SStefano 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]; 3527674ae819SStefano Zampini } 3528674ae819SStefano Zampini } 3529674ae819SStefano Zampini } 3530674ae819SStefano Zampini #else /* on missing GESVD */ 3531e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3532e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3533b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3534674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3535674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35369162d606SStefano 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)); 3537674ae819SStefano Zampini #else 35389162d606SStefano 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)); 3539674ae819SStefano Zampini #endif 3540984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3541674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3542984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3543e310c8b4SStefano Zampini k = temp_constraints; 3544e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3545674ae819SStefano Zampini j = 0; 3546e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 354745a1bb75SStefano Zampini valid_constraints = k-j; 3548911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3549984c4197SStefano Zampini #endif /* on missing GESVD */ 3550674ae819SStefano Zampini } 3551a773dcb8SStefano Zampini } 35529162d606SStefano Zampini /* update pointers information */ 35539162d606SStefano Zampini if (valid_constraints) { 35549162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 35559162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 35569162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 35579162d606SStefano Zampini /* set change_of_basis flag */ 355845a1bb75SStefano Zampini if (boolforchange) { 3559b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 35609162d606SStefano Zampini } 3561b3d85658SStefano Zampini total_counts_cc++; 356245a1bb75SStefano Zampini } 356345a1bb75SStefano Zampini } 3564984c4197SStefano Zampini /* free workspace */ 35658f1c130eSStefano Zampini if (!skip_lapack) { 3566984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3567984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3568984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3569984c4197SStefano Zampini #endif 3570984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3571984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3572984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3573984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3574984c4197SStefano Zampini #endif 3575984c4197SStefano Zampini } 3576984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3577984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3578984c4197SStefano Zampini } 3579984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3580cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3581cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3582cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3583cf5a6209SStefano Zampini } 3584cf5a6209SStefano Zampini if (n_ISForFaces) { 3585cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3586cf5a6209SStefano Zampini } 3587cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3588cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3589cf5a6209SStefano Zampini } 3590cf5a6209SStefano Zampini if (n_ISForEdges) { 3591cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3592cf5a6209SStefano Zampini } 3593cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 359408122e43SStefano Zampini } else { 359508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3596984c4197SStefano Zampini 359708122e43SStefano Zampini total_counts = 0; 359808122e43SStefano Zampini n_vertices = 0; 3599d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3600d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 360108122e43SStefano Zampini } 360208122e43SStefano Zampini max_constraints = 0; 36039162d606SStefano Zampini total_counts_cc = 0; 360408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 360508122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 36069162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 360708122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 360808122e43SStefano Zampini } 36099162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 36109162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 36119162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 36129162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 361374d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 36149162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 36159162d606SStefano Zampini total_counts_cc = 0; 36169162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 36179162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 36189162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 361908122e43SStefano Zampini } 362008122e43SStefano Zampini } 36219162d606SStefano Zampini #if 0 36229162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 36239162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36249162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 36259162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 36269162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 36279162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 36289162d606SStefano Zampini } 36299162d606SStefano Zampini printf("\n"); 36309162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 36319162d606SStefano Zampini } 36321b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 36338bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 36341b968477SStefano Zampini } 36351b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 36368bec7fa6SStefano 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]); 36371b968477SStefano Zampini } 363808122e43SStefano Zampini #endif 363908122e43SStefano Zampini 36408bec7fa6SStefano Zampini max_size_of_constraint = 0; 36419162d606SStefano 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]); 36429162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 364308122e43SStefano Zampini /* Change of basis */ 3644b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 364508122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 364608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 364708122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3648b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 364908122e43SStefano Zampini } 365008122e43SStefano Zampini } 365108122e43SStefano Zampini } 365208122e43SStefano Zampini } 3653984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 36544f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 365508122e43SStefano Zampini 36569162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 36579162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 36589162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 36599162d606SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for constraints indices %d != %d\n",constraints_idxs_ptr[total_counts_cc],i); 366008122e43SStefano Zampini } 3661674ae819SStefano Zampini 3662674ae819SStefano Zampini /* Create constraint matrix */ 3663674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 366416f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3665984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3666984c4197SStefano Zampini 3667984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3668a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3669a717540cSStefano Zampini qr_needed = PETSC_FALSE; 367074d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3671984c4197SStefano Zampini total_primal_vertices=0; 3672b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 36739162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36749162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3675984c4197SStefano Zampini if (size_of_constraint == 1) { 36769162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3677b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 367864efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 36799162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 36809162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3681a717540cSStefano Zampini } 3682b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 368374d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3684a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3685a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3686a717540cSStefano Zampini } 3687fa434743SStefano Zampini } else { 3688b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3689fa434743SStefano Zampini } 3690a717540cSStefano Zampini } 3691b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3692b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3693674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 369470022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3695b3d85658SStefano Zampini 36964f1b2e48SStefano 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); 36970e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 36980e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3699984c4197SStefano Zampini 3700984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 370174d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3702785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3703984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 370474d5cdf7SStefano Zampini 3705984c4197SStefano Zampini j = total_primal_vertices; 370674d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3707b3d85658SStefano Zampini cum = total_primal_vertices; 37089162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37094641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3710b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3711b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3712b3d85658SStefano Zampini cum++; 37139162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 371474d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 371574d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 371674d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 371774d5cdf7SStefano Zampini } 37189162d606SStefano Zampini j += constraints_n[i]; 3719674ae819SStefano Zampini } 3720674ae819SStefano Zampini } 3721674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3722674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3723088faed8SStefano Zampini 3724674ae819SStefano Zampini /* set values in constraint matrix */ 3725984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 37260e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3727674ae819SStefano Zampini } 3728984c4197SStefano Zampini total_counts = total_primal_vertices; 37299162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37304641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 37319162d606SStefano Zampini PetscInt *cols; 37329162d606SStefano Zampini 37339162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 37349162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 37359162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37369162d606SStefano Zampini PetscInt row = total_counts+k; 37379162d606SStefano Zampini PetscScalar *vals; 37389162d606SStefano Zampini 37399162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 37409162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 37419162d606SStefano Zampini } 37429162d606SStefano Zampini total_counts += constraints_n[i]; 3743674ae819SStefano Zampini } 3744674ae819SStefano Zampini } 3745674ae819SStefano Zampini /* assembling */ 3746674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3747674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3748088faed8SStefano Zampini 3749984c4197SStefano Zampini /* 375045a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3751984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3752984c4197SStefano Zampini */ 3753674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3754674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3755026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3756984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3757984c4197SStefano Zampini /* working stuff for GEQRF */ 375881d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3759984c4197SStefano Zampini PetscBLASInt lqr_work; 3760984c4197SStefano Zampini /* working stuff for UNGQR */ 3761984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3762984c4197SStefano Zampini PetscBLASInt lgqr_work; 3763984c4197SStefano Zampini /* working stuff for TRTRS */ 3764984c4197SStefano Zampini PetscScalar *trs_rhs; 37653f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3766984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3767984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3768984c4197SStefano Zampini PetscScalar *start_vals; 3769984c4197SStefano Zampini /* working stuff for values insertion */ 37704641a718SStefano Zampini PetscBT is_primal; 377164efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3772906d46d4SStefano Zampini /* matrix sizes */ 3773906d46d4SStefano Zampini PetscInt global_size,local_size; 3774906d46d4SStefano Zampini /* temporary change of basis */ 3775906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3776cf5a6209SStefano Zampini /* extra space for debugging */ 3777cf5a6209SStefano Zampini PetscScalar *dbg_work; 3778984c4197SStefano Zampini 3779906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3780906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 378116f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3782bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3783906d46d4SStefano Zampini /* nonzeros for local mat */ 3784bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3785bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 37869162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3787a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 37889162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3789a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 37909162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3791a717540cSStefano Zampini } else { 37929162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 37939162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3794a717540cSStefano Zampini } 3795a717540cSStefano Zampini } 3796a717540cSStefano Zampini } 3797906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3798bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3799a717540cSStefano Zampini /* Set initial identity in the matrix */ 3800bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3801906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3802a717540cSStefano Zampini } 3803a717540cSStefano Zampini 3804a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3805a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3806a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3807a717540cSStefano Zampini } 3808a717540cSStefano Zampini 3809a717540cSStefano Zampini 3810a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3811a717540cSStefano Zampini /* 3812a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3813a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3814a717540cSStefano Zampini 3815a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3816a717540cSStefano Zampini 3817a6b551f4SStefano 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) 3818a6b551f4SStefano Zampini 3819a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3820a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3821a717540cSStefano Zampini | ... | 3822a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3823a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3824a717540cSStefano Zampini 3825a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3826a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3827a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3828a6b551f4SStefano Zampini 3829a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3830a717540cSStefano Zampini */ 3831a717540cSStefano Zampini if (qr_needed) { 3832984c4197SStefano Zampini /* space to store Q */ 3833854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3834984c4197SStefano Zampini /* first we issue queries for optimal work */ 38353f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38363f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 38373f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3838984c4197SStefano Zampini lqr_work = -1; 38393f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3840984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3841984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3842785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3843984c4197SStefano Zampini lgqr_work = -1; 38443f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38453f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 38463f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 38473f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38483f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 38493f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3850984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3851984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3852785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3853984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3854785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3855984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3856785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3857a717540cSStefano Zampini /* allocating workspace for check */ 3858a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3859cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3860a717540cSStefano Zampini } 3861a717540cSStefano Zampini } 3862984c4197SStefano Zampini /* array to store whether a node is primal or not */ 38634641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3864473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 38650e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 386639e2fb2aSStefano Zampini if (i != total_primal_vertices) { 386739e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 38684641a718SStefano Zampini } 386939e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 387039e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 387139e2fb2aSStefano Zampini } 387239e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3873984c4197SStefano Zampini 3874a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 38759162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 38769162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 38774641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3878984c4197SStefano Zampini /* get constraint info */ 38799162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3880984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3881984c4197SStefano Zampini 3882984c4197SStefano Zampini if (pcbddc->dbg_flag) { 38839162d606SStefano 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); 3884674ae819SStefano Zampini } 3885984c4197SStefano Zampini 3886fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3887a717540cSStefano Zampini 3888a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3889a717540cSStefano Zampini if (pcbddc->dbg_flag) { 38909162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3891a717540cSStefano Zampini } 3892984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 38939162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3894984c4197SStefano Zampini 3895984c4197SStefano Zampini /* compute QR decomposition of constraints */ 38963f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 38973f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 38983f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3899674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39003f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3901984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3902674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3903984c4197SStefano Zampini 3904984c4197SStefano Zampini /* explictly compute R^-T */ 3905984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3906984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 39073f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39083f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 39093f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39103f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3911984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39123f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3913984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3914984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3915984c4197SStefano Zampini 3916a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 39173f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39183f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39193f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39203f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3921984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39223f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3923984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3924984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3925984c4197SStefano Zampini 3926984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3927984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3928984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 39293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39303f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39313f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39323f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39333f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 39343f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3935984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39369162d606SStefano 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)); 3937984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39389162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3939984c4197SStefano Zampini 3940984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 39419162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3942984c4197SStefano Zampini /* insert cols for primal dofs */ 3943984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3944984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 39459162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3946906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3947984c4197SStefano Zampini } 3948984c4197SStefano Zampini /* insert cols for dual dofs */ 3949984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 39509162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3951984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 39529162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3953906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3954984c4197SStefano Zampini j++; 3955674ae819SStefano Zampini } 3956674ae819SStefano Zampini } 3957984c4197SStefano Zampini 3958984c4197SStefano Zampini /* check change of basis */ 3959984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3960984c4197SStefano Zampini PetscInt ii,jj; 3961984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3962c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3963c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3964c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3965c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3966c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3967c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3968984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3969cf5a6209SStefano 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)); 3970984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3971984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3972984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3973cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3974cf5a6209SStefano 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; 3975674ae819SStefano Zampini } 3976674ae819SStefano Zampini } 3977984c4197SStefano Zampini if (!valid_qr) { 397822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3979984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3980984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3981cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3982cf5a6209SStefano 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])); 3983674ae819SStefano Zampini } 3984cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3985cf5a6209SStefano 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])); 3986984c4197SStefano Zampini } 3987984c4197SStefano Zampini } 3988984c4197SStefano Zampini } 3989674ae819SStefano Zampini } else { 399022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3991674ae819SStefano Zampini } 3992674ae819SStefano Zampini } 3993a717540cSStefano Zampini } else { /* simple transformation block */ 3994a717540cSStefano Zampini PetscInt row,col; 3995a6b551f4SStefano Zampini PetscScalar val,norm; 3996a6b551f4SStefano Zampini 3997a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39989162d606SStefano 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)); 3999a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 40009162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 40019162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4002bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 40039162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4004906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 40059162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4006a717540cSStefano Zampini } else { 4007a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 40089162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4009a717540cSStefano Zampini if (row != col) { 40109162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4011a717540cSStefano Zampini } else { 40129162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4013a717540cSStefano Zampini } 4014906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4015a717540cSStefano Zampini } 4016a717540cSStefano Zampini } 4017a717540cSStefano Zampini } 401898a51de6SStefano Zampini if (pcbddc->dbg_flag) { 401922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4020a717540cSStefano Zampini } 4021674ae819SStefano Zampini } 4022984c4197SStefano Zampini } else { 4023984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40249162d606SStefano 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); 4025674ae819SStefano Zampini } 4026674ae819SStefano Zampini } 4027674ae819SStefano Zampini } 4028a717540cSStefano Zampini 4029a717540cSStefano Zampini /* free workspace */ 4030a717540cSStefano Zampini if (qr_needed) { 4031984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4032cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4033984c4197SStefano Zampini } 4034984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4035984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4036984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4037984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4038984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4039674ae819SStefano Zampini } 4040a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4041906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4042906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4043906d46d4SStefano Zampini 4044906d46d4SStefano Zampini /* assembling of global change of variable */ 4045bbb9e6c6SStefano Zampini { 4046bbb9e6c6SStefano Zampini Mat tmat; 404716f15bc4SStefano Zampini PetscInt bs; 404816f15bc4SStefano Zampini 4049906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4050906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4051bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4052bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4053bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4054bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 405516f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 405616f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4057906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4058bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4059bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4060bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4061bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4062bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4063e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4064e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4065bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4066bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 4067906d46d4SStefano Zampini } 4068906d46d4SStefano Zampini /* check */ 4069906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4070906d46d4SStefano Zampini PetscReal error; 4071906d46d4SStefano Zampini Vec x,x_change; 4072906d46d4SStefano Zampini 4073906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4074906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4075906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4076906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4077e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4078e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4079bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4080e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4081e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4082906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4083906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4084906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4085906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4086bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4087906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4088906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4089906d46d4SStefano Zampini } 4090b96c3477SStefano Zampini 4091b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4092b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4093b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4094b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4095ac632422SStefano Zampini Mat S_new,tmat; 4096b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 4097b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 4098b087196eSStefano Zampini const PetscScalar *array; 4099b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 4100b087196eSStefano Zampini PetscInt i,n_V; 4101bbb9e6c6SStefano Zampini 4102bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 41036816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4104b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4105b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4106b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4107b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4108bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4109b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4110ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4111b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4112ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4113b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4114b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4115b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4116b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4117b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4118b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4119b087196eSStefano Zampini PetscScalar val; 4120b087196eSStefano Zampini PetscInt idx; 4121b087196eSStefano Zampini 4122b087196eSStefano Zampini idx = idxs_V[i]; 4123b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4124b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4125b087196eSStefano Zampini } 4126b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4127b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4128ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4129ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4130ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4131ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4132b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4133ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4134b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4135ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4136ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4137ac632422SStefano Zampini } 4138b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4139b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4140b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4141b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4142b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 4143b96c3477SStefano Zampini } 4144b96c3477SStefano Zampini } 4145906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 4146906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4147b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4148b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4149b9b85e73SStefano Zampini } 4150906d46d4SStefano Zampini 4151906d46d4SStefano Zampini /* set up change of basis context */ 4152906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4153906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4154906d46d4SStefano Zampini 4155906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4156906d46d4SStefano Zampini PetscInt global_size,local_size; 4157906d46d4SStefano Zampini 4158906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4159906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4160906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4161906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4162906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4163906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4164906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4165906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4166906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4167906d46d4SStefano Zampini } else { 4168906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4169906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4170906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4171906d46d4SStefano Zampini } 4172906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4173906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4174906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4175906d46d4SStefano Zampini } else { 4176906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4177906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4178906d46d4SStefano Zampini } 4179906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4180906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4181906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4182906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4183a8661815SStefano Zampini } else { 4184a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4185b9b85e73SStefano Zampini } 4186a717540cSStefano Zampini 41874f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 41884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 41894f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 41904f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4191019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4192019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4193019a44ceSStefano Zampini pcbddc->local_primal_size++; 4194019a44ceSStefano Zampini } 4195019a44ceSStefano Zampini 4196019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4197727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4198727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 41999f47a83aSStefano 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); 4200c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 42010e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 42029f47a83aSStefano 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); 4203727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4204727cdba6SStefano Zampini } 42050e6343abSStefano Zampini } 42060e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4207727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4208727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4209727cdba6SStefano Zampini 4210a717540cSStefano Zampini /* flush dbg viewer */ 4211b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4212b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4213b8ffe317SStefano Zampini } 4214a717540cSStefano Zampini 4215e310c8b4SStefano Zampini /* free workspace */ 4216a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 42174641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 421808122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 42199162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 42209162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 422108122e43SStefano Zampini } else { 42229162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 42239162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 42249162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 422508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 422608122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 42279162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 42289162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 422908122e43SStefano Zampini } 4230674ae819SStefano Zampini PetscFunctionReturn(0); 4231674ae819SStefano Zampini } 4232674ae819SStefano Zampini 4233674ae819SStefano Zampini #undef __FUNCT__ 4234674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4235674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4236674ae819SStefano Zampini { 4237674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4238674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4239674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 42407fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 4241674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4242674ae819SStefano Zampini 4243674ae819SStefano Zampini PetscFunctionBegin; 42448e61c736SStefano Zampini /* Reset previously computed graph */ 42458e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4246674ae819SStefano Zampini /* Init local Graph struct */ 42477fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 42483bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4249674ae819SStefano Zampini 4250575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 42515099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 42525099eff2SStefano 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); 4253575ad6abSStefano Zampini } 42549577ea80SStefano Zampini 4255674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4256d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 42574d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 42584d379d7bSStefano Zampini PetscInt nvtxs; 4259e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4260674ae819SStefano Zampini 42612fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42622fffb893SStefano Zampini if (flg_row) { 42634d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4264b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 42652fffb893SStefano Zampini } 42662fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4267674ae819SStefano Zampini } 42689b28b941SStefano Zampini if (pcbddc->dbg_flag) { 42699b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4270674ae819SStefano Zampini } 4271674ae819SStefano Zampini 427263602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 4273674ae819SStefano Zampini vertex_size = 1; 427463602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 427563602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 427695ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 427763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4278e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 427963602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4280674ae819SStefano Zampini } 428163602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 428263602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 428363602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4284674ae819SStefano Zampini } 428563602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 4286674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 428763602bcaSStefano Zampini } else { 428863602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 428963602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4290854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 429163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 429263602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 429363602bcaSStefano Zampini } 429463602bcaSStefano Zampini } 4295674ae819SStefano Zampini } 4296674ae819SStefano Zampini 4297674ae819SStefano Zampini /* Setup of Graph */ 4298785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4299e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4300785d1243SStefano Zampini } 4301785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4302e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4303785d1243SStefano Zampini } 430430368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 430530368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 430630368db7SStefano Zampini } 430730368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4308674ae819SStefano Zampini 43094f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 43104f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 43114f1b2e48SStefano Zampini PetscInt *local_subs; 43124f1b2e48SStefano Zampini 43134f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 43144f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 43154f1b2e48SStefano Zampini const PetscInt *idxs; 43164f1b2e48SStefano Zampini PetscInt nl,j; 43174f1b2e48SStefano Zampini 43184f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 43194f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43204f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 43214f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 43224f1b2e48SStefano Zampini } 43234f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43244f1b2e48SStefano Zampini } 43254f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 43264f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 43274f1b2e48SStefano Zampini } 43284f1b2e48SStefano Zampini 4329674ae819SStefano Zampini /* Graph's connected components analysis */ 4330674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4331674ae819SStefano Zampini 4332674ae819SStefano Zampini /* print some info to stdout */ 4333674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4334302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4335674ae819SStefano Zampini } 4336fb180af4SStefano Zampini 4337fb180af4SStefano Zampini /* mark topography has done */ 4338fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 4339674ae819SStefano Zampini PetscFunctionReturn(0); 4340674ae819SStefano Zampini } 4341674ae819SStefano Zampini 4342dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4343674ae819SStefano Zampini #undef __FUNCT__ 4344674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4345dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4346674ae819SStefano Zampini { 4347dc456d91SStefano Zampini PetscSF sf; 4348dc456d91SStefano Zampini PetscLayout map; 4349dc456d91SStefano Zampini const PetscInt *idxs; 4350dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4351dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4352dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4353dc456d91SStefano Zampini PetscMPIInt commsize; 4354674ae819SStefano Zampini PetscBool first_found; 4355674ae819SStefano Zampini PetscErrorCode ierr; 4356674ae819SStefano Zampini 4357674ae819SStefano Zampini PetscFunctionBegin; 4358dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4359dc456d91SStefano Zampini if (subset_mult) { 4360dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4361dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4362dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4363674ae819SStefano Zampini } 4364dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4365dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4366dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4367dc456d91SStefano Zampini for (i=0;i<n;i++) { 4368dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4369dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4370674ae819SStefano Zampini } 4371dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4372dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4373dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4374dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4375dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4376dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4377dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4378dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4379dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4380dc456d91SStefano Zampini 4381dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4382dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4383dc456d91SStefano Zampini if (subset_mult) { 4384dc456d91SStefano Zampini const PetscInt* idxs_mult; 4385dc456d91SStefano Zampini 4386dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4387dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4388dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4389674ae819SStefano Zampini } else { 4390dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4391674ae819SStefano Zampini } 4392dc456d91SStefano Zampini /* local size of new subset */ 4393dc456d91SStefano Zampini n_n = 0; 4394dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4395dc456d91SStefano Zampini 4396dc456d91SStefano Zampini /* global indexes in layout */ 4397dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4398dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4399dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4400dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4401dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4402dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4403dc456d91SStefano Zampini 4404dc456d91SStefano Zampini /* reduce from leaves to roots */ 4405dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 440664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 440764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4408dc456d91SStefano Zampini 4409dc456d91SStefano Zampini /* count indexes in local part of layout */ 4410674ae819SStefano Zampini nlocals = 0; 4411674ae819SStefano Zampini first_index = -1; 4412674ae819SStefano Zampini first_found = PETSC_FALSE; 4413dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4414dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4415674ae819SStefano Zampini first_found = PETSC_TRUE; 4416674ae819SStefano Zampini first_index = i; 4417674ae819SStefano Zampini } 4418dc456d91SStefano Zampini nlocals += root_data[i]; 4419674ae819SStefano Zampini } 4420dc456d91SStefano Zampini 4421dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 44225fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4423dc456d91SStefano Zampini start = 0; 442464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44255fa240b1SStefano Zampini #else 442664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44275fa240b1SStefano Zampini start = start-nlocals; 44285fa240b1SStefano Zampini #endif 44295fa240b1SStefano Zampini 4430dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4431dc456d91SStefano Zampini *N_n = start + nlocals; 4432dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4433dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4434674ae819SStefano Zampini } 44355fa240b1SStefano Zampini 44365fa240b1SStefano Zampini /* adapt root data with cumulative */ 4437674ae819SStefano Zampini if (first_found) { 4438dc456d91SStefano Zampini PetscInt old_index; 4439dc456d91SStefano Zampini 4440dc456d91SStefano Zampini root_data[first_index] += start; 4441674ae819SStefano Zampini old_index = first_index; 4442dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4443dc456d91SStefano Zampini if (root_data[i]) { 4444dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4445674ae819SStefano Zampini old_index = i; 4446674ae819SStefano Zampini } 4447674ae819SStefano Zampini } 4448674ae819SStefano Zampini } 4449dc456d91SStefano Zampini 4450dc456d91SStefano Zampini /* from roots to leaves */ 4451dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4452dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4453dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4454dc456d91SStefano Zampini 4455dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4456dc456d91SStefano Zampini if (subset_mult) { 4457dc456d91SStefano Zampini const PetscInt* idxs_mult; 4458dc456d91SStefano Zampini PetscInt cum; 4459dc456d91SStefano Zampini 4460dc456d91SStefano Zampini cum = 0; 4461dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4462dc456d91SStefano Zampini for (i=0;i<n;i++) { 4463dc456d91SStefano Zampini PetscInt j; 4464dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4465674ae819SStefano Zampini } 4466dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4467674ae819SStefano Zampini } else { 4468dc456d91SStefano Zampini for (i=0;i<n;i++) { 4469dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4470674ae819SStefano Zampini } 4471674ae819SStefano Zampini } 4472dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4473dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4474674ae819SStefano Zampini PetscFunctionReturn(0); 4475674ae819SStefano Zampini } 44769a7d3425SStefano Zampini 44779a7d3425SStefano Zampini #undef __FUNCT__ 44789a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 44799a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 44809a7d3425SStefano Zampini { 44819a7d3425SStefano Zampini PetscInt i,j; 44829a7d3425SStefano Zampini PetscScalar *alphas; 44839a7d3425SStefano Zampini PetscErrorCode ierr; 44849a7d3425SStefano Zampini 44859a7d3425SStefano Zampini PetscFunctionBegin; 44869a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4487785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 44889a7d3425SStefano Zampini for (i=0;i<n;i++) { 44899a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 44909a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 44919a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 44929a7d3425SStefano Zampini } 44939a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 44949a7d3425SStefano Zampini PetscFunctionReturn(0); 44959a7d3425SStefano Zampini } 44969a7d3425SStefano Zampini 4497e7931f94SStefano Zampini #undef __FUNCT__ 449870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4499b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4500e7931f94SStefano Zampini { 450152e5ac9dSStefano Zampini IS ranks_send_to; 4502e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4503e7931f94SStefano Zampini PetscMPIInt size,rank,color; 450452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 450552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 45063837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 45072b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4508e7931f94SStefano Zampini PetscSubcomm subcomm; 450952e5ac9dSStefano Zampini PetscErrorCode ierr; 4510a57a6d2fSStefano Zampini 4511e7931f94SStefano Zampini PetscFunctionBegin; 45122b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 45132b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 45142b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4515e7931f94SStefano Zampini 4516e7931f94SStefano Zampini /* Get info on mapping */ 45173bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 45183bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4519e7931f94SStefano Zampini 4520e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4521785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4522e7931f94SStefano Zampini xadj[0] = 0; 4523e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4524785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4525785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4526e7931f94SStefano Zampini 45272b510759SStefano Zampini if (threshold) { 4528d023bfaeSStefano Zampini PetscInt xadj_count = 0; 45292b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4530d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4531d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4532d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4533d023bfaeSStefano Zampini xadj_count++; 4534e7931f94SStefano Zampini } 4535e7931f94SStefano Zampini } 4536d023bfaeSStefano Zampini xadj[1] = xadj_count; 4537c8587f34SStefano Zampini } else { 4538e7931f94SStefano Zampini if (xadj[1]) { 4539e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4540e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4541c8587f34SStefano Zampini } 4542e7931f94SStefano Zampini } 45433bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4544e7931f94SStefano Zampini if (use_square) { 4545e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4546e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4547e7931f94SStefano Zampini } 4548e7931f94SStefano Zampini } 4549e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4550e7931f94SStefano Zampini 45513837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4552e7931f94SStefano Zampini 4553e7931f94SStefano Zampini /* 4554e7931f94SStefano Zampini Restrict work on active processes only. 4555e7931f94SStefano Zampini */ 4556e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4557e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4558e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 45592b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4560d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4561e7931f94SStefano Zampini if (color) { 4562e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4563e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4564e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4565c8587f34SStefano Zampini } else { 456652e5ac9dSStefano Zampini Mat subdomain_adj; 456752e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 456852e5ac9dSStefano Zampini MatPartitioning partitioner; 456952e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 457052e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4571b0c7d250SStefano Zampini PetscBool aggregate; 4572b0c7d250SStefano Zampini 4573306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4574785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4575e7931f94SStefano Zampini prank = rank; 4576306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 45778002ef2cSStefano Zampini /* 4578e7931f94SStefano Zampini for (i=0;i<size;i++) { 4579e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4580c8587f34SStefano Zampini } 45818002ef2cSStefano Zampini */ 4582e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4583e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4584c8587f34SStefano Zampini } 4585e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4586b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4587b0c7d250SStefano Zampini if (aggregate) { 4588b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4589b0c7d250SStefano Zampini PetscMPIInt nrank; 4590b0c7d250SStefano Zampini PetscScalar *vals; 4591b0c7d250SStefano Zampini 4592b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4593b0c7d250SStefano Zampini lrows = 0; 4594b0c7d250SStefano Zampini if (nrank<redprocs) { 4595b0c7d250SStefano Zampini lrows = size/redprocs; 4596b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4597b0c7d250SStefano Zampini } 45985fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4599b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4600b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4601b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4602b0c7d250SStefano Zampini row = nrank; 4603b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4604b0c7d250SStefano Zampini cols = adjncy; 4605b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4606b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4607b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4608b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4609b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 461052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 461152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 461252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4613b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4614b0c7d250SStefano Zampini } else { 4615306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4616b0c7d250SStefano Zampini } 461722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4618e7931f94SStefano Zampini 4619e7931f94SStefano Zampini /* Partition */ 4620306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4621e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4622e7931f94SStefano Zampini if (use_vwgt) { 46233837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4624e7931f94SStefano Zampini v_wgt[0] = local_size; 4625e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4626c8587f34SStefano Zampini } 462728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 462828143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4629e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4630e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 463122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4632e7931f94SStefano Zampini 463352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 463452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 463552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 463652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4637b0c7d250SStefano Zampini if (!redprocs) { 4638b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 463928143c3dSStefano Zampini } else { 4640b0c7d250SStefano Zampini PetscInt idxs[1]; 4641b0c7d250SStefano Zampini PetscMPIInt tag; 4642b0c7d250SStefano Zampini MPI_Request *reqs; 4643b0c7d250SStefano Zampini 4644b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4645b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4646b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4647b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 464828143c3dSStefano Zampini } 4649b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4650b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4651b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4652b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4653e7931f94SStefano Zampini } 465452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4655e7931f94SStefano Zampini /* clean up */ 4656e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 465752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4658e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4659e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4660e7931f94SStefano Zampini } 4661e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4662e7931f94SStefano Zampini 4663e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4664e7931f94SStefano Zampini i = 1; 4665e7931f94SStefano Zampini if (color) i=0; 4666e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4667e7931f94SStefano Zampini /* get back IS */ 4668e7931f94SStefano Zampini *is_sends = ranks_send_to; 4669e7931f94SStefano Zampini PetscFunctionReturn(0); 4670e7931f94SStefano Zampini } 4671e7931f94SStefano Zampini 4672e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4673e7931f94SStefano Zampini 4674e7931f94SStefano Zampini #undef __FUNCT__ 4675e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 467653a05cb3SStefano 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[]) 4677e7931f94SStefano Zampini { 467870cf5478SStefano Zampini Mat local_mat; 4679e7931f94SStefano Zampini IS is_sends_internal; 46809d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 468128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 46829d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4683e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4684e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4685e7931f94SStefano Zampini const PetscInt* is_indices; 4686e7931f94SStefano Zampini MatType new_local_type; 4687e7931f94SStefano Zampini /* buffers */ 4688e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 468928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 46909d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4691e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4692e7931f94SStefano Zampini /* MPI */ 469328143c3dSStefano Zampini MPI_Comm comm,comm_n; 469428143c3dSStefano Zampini PetscSubcomm subcomm; 4695e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 469628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 469728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 469828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 469928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 470028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4701e7931f94SStefano Zampini PetscErrorCode ierr; 4702e7931f94SStefano Zampini 4703e7931f94SStefano Zampini PetscFunctionBegin; 470428143c3dSStefano Zampini /* TODO: add missing checks */ 470528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 470628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 470728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 470828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4709e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 471028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4711e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4712e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4713e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4714e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4715e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 471628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 471770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 471870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 471928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 472070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 472170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 472270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 472370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 472470cf5478SStefano Zampini } 4725e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4726e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4727e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4728e7931f94SStefano Zampini if (!is_sends) { 472928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4730b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4731c8587f34SStefano Zampini } else { 4732e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4733e7931f94SStefano Zampini is_sends_internal = is_sends; 4734c8587f34SStefano Zampini } 4735e7931f94SStefano Zampini 4736e7931f94SStefano Zampini /* get comm */ 4737a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4738e7931f94SStefano Zampini 4739e7931f94SStefano Zampini /* compute number of sends */ 4740e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4741e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4742e7931f94SStefano Zampini 4743e7931f94SStefano Zampini /* compute number of receives */ 4744e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4745785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4746e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4747e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4748e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4749e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4750e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4751e7931f94SStefano Zampini 475228143c3dSStefano Zampini /* restrict comm if requested */ 475328143c3dSStefano Zampini subcomm = 0; 475428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 475528143c3dSStefano Zampini if (restrict_comm) { 4756779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4757779c1cceSStefano Zampini 475828143c3dSStefano Zampini color = 0; 475953a05cb3SStefano Zampini if (restrict_full) { 476053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 476153a05cb3SStefano Zampini } else { 476253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 476353a05cb3SStefano Zampini } 476428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 476528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 476628143c3dSStefano Zampini /* check if reuse has been requested */ 476728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 476828143c3dSStefano Zampini if (*mat_n) { 476928143c3dSStefano Zampini PetscMPIInt subcommsize2; 477028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 477128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 477228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 477328143c3dSStefano Zampini } else { 477428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 477528143c3dSStefano Zampini } 477628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4777779c1cceSStefano Zampini PetscMPIInt rank; 4778779c1cceSStefano Zampini 4779779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 478028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 478128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 478228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4783306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 478428143c3dSStefano Zampini } 478528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 478628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 478728143c3dSStefano Zampini } else { 478828143c3dSStefano Zampini comm_n = comm; 478928143c3dSStefano Zampini } 479028143c3dSStefano Zampini 4791e7931f94SStefano Zampini /* prepare send/receive buffers */ 4792785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4793e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4794785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4795e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 479628143c3dSStefano Zampini if (nis) { 4797854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 479828143c3dSStefano Zampini } 4799e7931f94SStefano Zampini 480028143c3dSStefano Zampini /* Get data from local matrices */ 4801e7931f94SStefano Zampini if (!isdense) { 4802a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4803e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4804e7931f94SStefano Zampini /* 4805e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4806e7931f94SStefano Zampini send_buffer_idxs should contain: 4807e7931f94SStefano Zampini - MatType_PRIVATE type 4808e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4809e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4810e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4811e7931f94SStefano Zampini */ 4812e7931f94SStefano Zampini } else { 4813e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 48143bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4815854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4816e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4817e7931f94SStefano Zampini send_buffer_idxs[1] = i; 48183bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4819e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 48203bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4821e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4822e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4823e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4824e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4825c8587f34SStefano Zampini } 4826c8587f34SStefano Zampini } 4827e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 482828143c3dSStefano Zampini /* additional is (if any) */ 482928143c3dSStefano Zampini if (nis) { 483028143c3dSStefano Zampini PetscMPIInt psum; 483128143c3dSStefano Zampini PetscInt j; 483228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 483328143c3dSStefano Zampini PetscInt plen; 483428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 483528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 483628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 483728143c3dSStefano Zampini } 4838854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 483928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 484028143c3dSStefano Zampini PetscInt plen; 484128143c3dSStefano Zampini const PetscInt *is_array_idxs; 484228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 484328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 484428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 484528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 484628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 484728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 484828143c3dSStefano Zampini } 484928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 485028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 485128143c3dSStefano Zampini } 485228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 485328143c3dSStefano Zampini } 485428143c3dSStefano Zampini 4855e7931f94SStefano Zampini buf_size_idxs = 0; 4856e7931f94SStefano Zampini buf_size_vals = 0; 485728143c3dSStefano Zampini buf_size_idxs_is = 0; 4858e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4859e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4860e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 486128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4862e7931f94SStefano Zampini } 4863785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4864785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 486595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4866e7931f94SStefano Zampini 4867e7931f94SStefano Zampini /* get new tags for clean communications */ 4868e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4869e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 487028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4871e7931f94SStefano Zampini 4872e7931f94SStefano Zampini /* allocate for requests */ 4873785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4874785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 487595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4876785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4877785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 487895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4879e7931f94SStefano Zampini 4880e7931f94SStefano Zampini /* communications */ 4881e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4882e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 488328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4884e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4885e7931f94SStefano Zampini source_dest = onodes[i]; 4886e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4887e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4888e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4889e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 489028143c3dSStefano Zampini if (nis) { 489128143c3dSStefano 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); 489228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 489328143c3dSStefano Zampini } 4894e7931f94SStefano Zampini } 4895e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4896e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4897e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4898e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 489928143c3dSStefano Zampini if (nis) { 490028143c3dSStefano 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); 490128143c3dSStefano Zampini } 4902e7931f94SStefano Zampini } 4903e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4904e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4905e7931f94SStefano Zampini 4906e7931f94SStefano Zampini /* assemble new l2g map */ 4907e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4908e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49099d30be91SStefano Zampini new_local_rows = 0; 4910e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49119d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4912e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4913e7931f94SStefano Zampini } 49149d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4915e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49169d30be91SStefano Zampini new_local_rows = 0; 4917e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49189d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 49199d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4920e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4921e7931f94SStefano Zampini } 49229d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 49239d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4924e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4925e7931f94SStefano Zampini 4926e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4927e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4928e7931f94SStefano 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) */ 4929e7931f94SStefano Zampini if (n_recvs) { 493028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4931e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4932e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4933e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4934e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4935e7931f94SStefano Zampini break; 4936e7931f94SStefano Zampini } 4937e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4938e7931f94SStefano Zampini } 4939e7931f94SStefano Zampini switch (new_local_type_private) { 494028143c3dSStefano Zampini case MATDENSE_PRIVATE: 494128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4942e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4943e7931f94SStefano Zampini bs = 1; 494428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 494528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 494628143c3dSStefano Zampini bs = 1; 494728143c3dSStefano Zampini } 4948e7931f94SStefano Zampini break; 4949e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4950e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4951e7931f94SStefano Zampini bs = 1; 4952e7931f94SStefano Zampini break; 4953e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4954e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4955e7931f94SStefano Zampini break; 4956e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4957e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4958e7931f94SStefano Zampini break; 4959e7931f94SStefano Zampini default: 49609d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4961e7931f94SStefano Zampini break; 4962e7931f94SStefano Zampini } 496328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 496428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 496528143c3dSStefano Zampini bs = 1; 4966e7931f94SStefano Zampini } 4967e7931f94SStefano Zampini 496870cf5478SStefano Zampini /* create MATIS object if needed */ 496970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 4970e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4971e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 497270cf5478SStefano Zampini } else { 497370cf5478SStefano Zampini /* it also destroys the local matrices */ 497470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 497570cf5478SStefano Zampini } 497670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 4977e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 49789d30be91SStefano Zampini 49799d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 49809d30be91SStefano Zampini 49819d30be91SStefano Zampini /* Global to local map of received indices */ 49829d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 49839d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 49849d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 49859d30be91SStefano Zampini 49869d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 49879d30be91SStefano Zampini buf_size_idxs = 0; 49889d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 49899d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 49909d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 49919d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 49929d30be91SStefano Zampini } 49939d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 49949d30be91SStefano Zampini 49959d30be91SStefano Zampini /* set preallocation */ 49969d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 49979d30be91SStefano Zampini if (!newisdense) { 49989d30be91SStefano Zampini PetscInt *new_local_nnz=0; 49999d30be91SStefano Zampini 50009d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 50019d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 50029d30be91SStefano Zampini if (n_recvs) { 50039d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 50049d30be91SStefano Zampini } 50059d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50069d30be91SStefano Zampini PetscInt j; 50079d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 50089d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 50099d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 50109d30be91SStefano Zampini } 50119d30be91SStefano Zampini } else { 50129d30be91SStefano Zampini /* TODO */ 50139d30be91SStefano Zampini } 50149d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 50159d30be91SStefano Zampini } 50169d30be91SStefano Zampini if (new_local_nnz) { 50179d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 50189d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 50199d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 50209d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50219d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 50229d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50239d30be91SStefano Zampini } else { 50249d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50259d30be91SStefano Zampini } 50269d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 50279d30be91SStefano Zampini } else { 50289d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50299d30be91SStefano Zampini } 5030e7931f94SStefano Zampini 5031e7931f94SStefano Zampini /* set values */ 5032e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 50339d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5034e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5035e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5036e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 50379d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5038e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5039e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5040e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 504128143c3dSStefano Zampini } else { 504228143c3dSStefano Zampini /* TODO */ 5043e7931f94SStefano Zampini } 5044e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5045e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5046e7931f94SStefano Zampini } 5047e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5048e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 504970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 505070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 50519d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 50529d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5053e7931f94SStefano Zampini 5054dfd14d43SStefano Zampini #if 0 505528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5056e7931f94SStefano Zampini Vec lvec,rvec; 5057e7931f94SStefano Zampini PetscReal infty_error; 5058e7931f94SStefano Zampini 50592a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5060e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5061e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5062e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 506370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5064e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5065e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5066e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5067e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5068e7931f94SStefano Zampini } 506928143c3dSStefano Zampini #endif 5070e7931f94SStefano Zampini 507128143c3dSStefano Zampini /* assemble new additional is (if any) */ 507228143c3dSStefano Zampini if (nis) { 507328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 507428143c3dSStefano Zampini 507528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5076854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 507728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 507828143c3dSStefano Zampini psum = 0; 507928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 508028143c3dSStefano Zampini for (j=0;j<nis;j++) { 508128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 508228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 508328143c3dSStefano Zampini psum += plen; 508428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 508528143c3dSStefano Zampini } 508628143c3dSStefano Zampini } 5087854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5088854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 508928143c3dSStefano Zampini for (i=1;i<nis;i++) { 509028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 509128143c3dSStefano Zampini } 509228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 509328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 509428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 509528143c3dSStefano Zampini for (j=0;j<nis;j++) { 509628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 509728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 509828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 509928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 510028143c3dSStefano Zampini } 510128143c3dSStefano Zampini } 510228143c3dSStefano Zampini for (i=0;i<nis;i++) { 510328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 510428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 510528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 510628143c3dSStefano Zampini } 510728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 510828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 510928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 511028143c3dSStefano Zampini } 5111e7931f94SStefano Zampini /* free workspace */ 511228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5113e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5114e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5115e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5116e7931f94SStefano Zampini if (isdense) { 5117e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5118e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5119e7931f94SStefano Zampini } else { 5120e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5121e7931f94SStefano Zampini } 512228143c3dSStefano Zampini if (nis) { 512328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 512428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 512528143c3dSStefano Zampini } 5126e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5127e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 512828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5129e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5130e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 513128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5132e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5133e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5134e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5135e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5136e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 513728143c3dSStefano Zampini if (nis) { 513828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 513928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 514028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 514128143c3dSStefano Zampini } 514228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 514328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 514428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 514528143c3dSStefano Zampini for (i=0;i<nis;i++) { 514628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 514728143c3dSStefano Zampini } 514853a05cb3SStefano Zampini *mat_n = NULL; 514928143c3dSStefano Zampini } 5150e7931f94SStefano Zampini PetscFunctionReturn(0); 5151e7931f94SStefano Zampini } 5152a57a6d2fSStefano Zampini 515312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5154af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 515512edc857SStefano Zampini 5156c8587f34SStefano Zampini #undef __FUNCT__ 5157c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5158c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5159c8587f34SStefano Zampini { 5160c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5161c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 516220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 51639881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 516420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 51656e683305SStefano Zampini IS coarse_is,*isarray; 51666e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 516730368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5168f9eb5b7dSStefano Zampini PC pc_temp; 5169c8587f34SStefano Zampini PCType coarse_pc_type; 5170c8587f34SStefano Zampini KSPType coarse_ksp_type; 5171f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 51724f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 51736e683305SStefano Zampini Mat t_coarse_mat_is; 51746e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 51756e683305SStefano Zampini PetscMPIInt all_procs; 517674e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 517768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 517822bc73bbSStefano Zampini PetscScalar *array; 51799881197aSStefano Zampini PetscErrorCode ierr; 5180fdc09c96SStefano Zampini 5181c8587f34SStefano Zampini PetscFunctionBegin; 5182c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 518368457ee5SStefano 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 */ 5184fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 51855a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5186fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5187f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5188f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5189f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5190fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 519151bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 519251bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5193dc4bcba2SStefano Zampini PC pc; 5194dc4bcba2SStefano Zampini PetscBool isbddc; 5195dc4bcba2SStefano Zampini 5196dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5197dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5198dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5199dc4bcba2SStefano Zampini if (isbddc) { 520063c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 520163c961adSStefano Zampini } else { 5202727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 520363c961adSStefano Zampini } 5204fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5205fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5206fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5207f4ddd8eeSStefano Zampini } 5208fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5209fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5210f4ddd8eeSStefano Zampini } 521170cf5478SStefano Zampini /* reset any subassembling information */ 521270cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 52136e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 52146e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5215fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5216f4ddd8eeSStefano Zampini } 5217c8587f34SStefano Zampini 52186e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 52192b510759SStefano Zampini im_active = !!(pcis->n); 52202b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 52216e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 52226e683305SStefano Zampini void_procs = all_procs-active_procs; 52236e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 522474e2c79eSStefano Zampini redist = PETSC_FALSE; 522522bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 52266e683305SStefano Zampini csin_ml = PETSC_TRUE; 52276e683305SStefano Zampini ncoarse_ml = void_procs; 5228779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5229779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 52306e683305SStefano Zampini csin_ds = PETSC_TRUE; 523118a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 523218a45a71SStefano Zampini redist = PETSC_TRUE; 523318a45a71SStefano Zampini } else { 52346e683305SStefano Zampini csin_ds = PETSC_TRUE; 5235779c1cceSStefano Zampini ncoarse_ds = active_procs; 5236779c1cceSStefano Zampini redist = PETSC_TRUE; 523718a45a71SStefano Zampini } 52386e683305SStefano Zampini } else { 52396e683305SStefano Zampini csin_ml = PETSC_FALSE; 52406e683305SStefano Zampini ncoarse_ml = all_procs; 52416e683305SStefano Zampini if (void_procs) { 52426e683305SStefano Zampini csin_ds = PETSC_TRUE; 52436e683305SStefano Zampini ncoarse_ds = void_procs; 52446e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 52456e683305SStefano Zampini } else { 5246779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 524774e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 524874e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 524974e2c79eSStefano Zampini redist = PETSC_TRUE; 525074e2c79eSStefano Zampini } else { 52516e683305SStefano Zampini csin_ds = PETSC_FALSE; 52526e683305SStefano Zampini ncoarse_ds = all_procs; 52536e683305SStefano Zampini } 52546e683305SStefano Zampini } 525574e2c79eSStefano Zampini } 52566e683305SStefano Zampini 52576e683305SStefano Zampini /* 52586e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 52596e683305SStefano Zampini - we have not exceeded the number of levels requested 52606e683305SStefano Zampini - we can actually subassemble the active processes 52616e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 52626e683305SStefano Zampini */ 52636e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 52646e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 52656e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 52666e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 52676e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5268f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 52692b510759SStefano Zampini } else { 5270f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5271c8587f34SStefano Zampini } 5272c8587f34SStefano Zampini } 52736e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 52746e683305SStefano Zampini if (multilevel_allowed) { 52756e683305SStefano Zampini ncoarse = ncoarse_ml; 52766e683305SStefano Zampini csin = csin_ml; 527758da7f69SStefano Zampini redist = PETSC_FALSE; 52786e683305SStefano Zampini } else { 52796e683305SStefano Zampini ncoarse = ncoarse_ds; 52806e683305SStefano Zampini csin = csin_ds; 52816e683305SStefano Zampini } 5282e7931f94SStefano Zampini 5283abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5284abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5285abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5286abbbba34SStefano Zampini 5287abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 528822bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 528922bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 529022bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 529122bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5292e176bc59SStefano 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); 52936e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 52946e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 52956e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5296abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5297abbbba34SStefano Zampini 52986e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 529930368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 53006e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 53016e683305SStefano Zampini const PetscInt *idxs; 53026e683305SStefano Zampini ISLocalToGlobalMapping tmap; 53036e683305SStefano Zampini 53046e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 53050be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 53066e683305SStefano Zampini /* allocate space for temporary storage */ 5307854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5308854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 53096e683305SStefano Zampini /* allocate for IS array */ 53106e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 53116e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 531230368db7SStefano Zampini nisvert = 0; 531330368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 531430368db7SStefano Zampini nisvert = 1; 531530368db7SStefano Zampini } 531630368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5317854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 53186e683305SStefano Zampini /* dofs splitting */ 53196e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 53206e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 53216e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 53226e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53236e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53246e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53256e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 532630368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 53276e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 53286e683305SStefano Zampini } 53296e683305SStefano Zampini /* neumann boundaries */ 53306e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 53316e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 53326e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 53336e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53346e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53356e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53366e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 533730368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 53386e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 53396e683305SStefano Zampini } 534030368db7SStefano Zampini /* primal vertices (benign) */ 534130368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 534230368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 534330368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 534430368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 534530368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 534630368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 534730368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 534830368db7SStefano Zampini } 53496e683305SStefano Zampini /* free memory */ 53506e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 53516e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 53526e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 53536e683305SStefano Zampini } else { 53546e683305SStefano Zampini nis = 0; 53556e683305SStefano Zampini nisdofs = 0; 53566e683305SStefano Zampini nisneu = 0; 535730368db7SStefano Zampini nisvert = 0; 53586e683305SStefano Zampini isarray = NULL; 53596e683305SStefano Zampini } 53606e683305SStefano Zampini /* destroy no longer needed map */ 53616e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 53626e683305SStefano Zampini 53636e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 53646e683305SStefano Zampini coarse_mat_is = NULL; 53656e683305SStefano Zampini if (csin) { 53666e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 536774e2c79eSStefano Zampini if (redist) { 536874e2c79eSStefano Zampini PetscMPIInt rank; 5369779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 537074e2c79eSStefano Zampini 537174e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 537258da7f69SStefano Zampini spc = active_procs/ncoarse; 537358da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5374779c1cceSStefano Zampini if (im_active) { 5375779c1cceSStefano Zampini destsize = 1; 537674e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 537774e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 537874e2c79eSStefano Zampini } else { 537974e2c79eSStefano Zampini dest[0] = rank/(spc+1); 538074e2c79eSStefano Zampini } 538174e2c79eSStefano Zampini } else { 5382779c1cceSStefano Zampini destsize = 0; 53836e683305SStefano Zampini } 5384779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5385779c1cceSStefano Zampini } else if (csin_type_simple) { 53866e683305SStefano Zampini PetscMPIInt rank; 53876e683305SStefano Zampini PetscInt issize,isidx; 5388779c1cceSStefano Zampini 53896e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 53906e683305SStefano Zampini if (im_active) { 53916e683305SStefano Zampini issize = 1; 53926e683305SStefano Zampini isidx = (PetscInt)rank; 53936e683305SStefano Zampini } else { 53946e683305SStefano Zampini issize = 0; 53956e683305SStefano Zampini isidx = -1; 53966e683305SStefano Zampini } 53976e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5398779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5399b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54006e683305SStefano Zampini } 5401779c1cceSStefano Zampini 5402779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5403779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5404779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5405779c1cceSStefano Zampini PetscInt *coarse_candidates; 5406779c1cceSStefano Zampini const PetscInt* tisindices; 5407779c1cceSStefano Zampini 5408779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5409779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5410779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5411779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5412779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5413779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5414779c1cceSStefano Zampini } 5415779c1cceSStefano Zampini } 5416779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5417779c1cceSStefano Zampini 5418779c1cceSStefano Zampini 54196e683305SStefano Zampini if (pcbddc->dbg_flag) { 54206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54216e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 54226e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5424779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 54256e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 54266e683305SStefano Zampini } 54276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 54286e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54296e683305SStefano Zampini } 54306e683305SStefano Zampini /* shift the pattern on coarse candidates */ 54316e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 54326e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5433854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 54346e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 54356e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 54366e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 54376e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 54386e683305SStefano Zampini } 54396e683305SStefano Zampini if (pcbddc->dbg_flag) { 54406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 54426e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54436e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54446e683305SStefano Zampini } 5445779c1cceSStefano Zampini } 54466e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 544753a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 544853a05cb3SStefano 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); 544953a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 545053a05cb3SStefano 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); 545153a05cb3SStefano Zampini } 54526e683305SStefano Zampini } else { 54536e683305SStefano Zampini if (pcbddc->dbg_flag) { 54546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 54566e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54576e683305SStefano Zampini } 54586e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 54596e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 54606e683305SStefano Zampini } 54616e683305SStefano Zampini 54626e683305SStefano Zampini /* create local to global scatters for coarse problem */ 546368457ee5SStefano Zampini if (compute_vecs) { 54646e683305SStefano Zampini PetscInt lrows; 54656e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 54666e683305SStefano Zampini if (coarse_mat_is) { 54676e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 54686e683305SStefano Zampini } else { 54696e683305SStefano Zampini lrows = 0; 54706e683305SStefano Zampini } 54716e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 54726e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 54736e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 54746e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 54756e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 54766e683305SStefano Zampini } 54776e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 54786e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5479c8587f34SStefano Zampini 5480f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5481f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5482f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5483f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5484f9eb5b7dSStefano Zampini } else { 5485f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5486f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5487c8587f34SStefano Zampini } 5488c8587f34SStefano Zampini 54896e683305SStefano Zampini /* print some info if requested */ 54906e683305SStefano Zampini if (pcbddc->dbg_flag) { 54916e683305SStefano Zampini if (!multilevel_allowed) { 54926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54936e683305SStefano Zampini if (multilevel_requested) { 54946e683305SStefano 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); 54956e683305SStefano Zampini } else if (pcbddc->max_levels) { 54966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 54976e683305SStefano Zampini } 54986e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54996e683305SStefano Zampini } 55006e683305SStefano Zampini } 55016e683305SStefano Zampini 5502f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 55036e683305SStefano Zampini if (coarse_mat_is) { 55046e683305SStefano Zampini MatReuse coarse_mat_reuse; 55056a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 55066e683305SStefano Zampini if (pcbddc->dbg_flag) { 55076e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 55086e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 55096e683305SStefano Zampini } 5510f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5511312be037SStefano Zampini char prefix[256],str_level[16]; 5512e604994aSStefano Zampini size_t len; 55136e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5514422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5515c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5516f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 55175f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5518c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 55196e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5520c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5521c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5522e604994aSStefano Zampini /* prefix */ 5523e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5524e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5525e604994aSStefano Zampini if (!pcbddc->current_level) { 5526e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5527e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5528c8587f34SStefano Zampini } else { 5529e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5530312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5531312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 553234d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5533312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5534e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5535e604994aSStefano Zampini } 5536e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 55373e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 55383e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 55393e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 55403e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5541f9eb5b7dSStefano Zampini /* allow user customization */ 5542f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 55433e3c6dadSStefano Zampini } 55443e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 554551bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 55463e3c6dadSStefano Zampini if (nisdofs) { 55473e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 55483e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 55493e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 55503e3c6dadSStefano Zampini } 55513e3c6dadSStefano Zampini } 55523e3c6dadSStefano Zampini if (nisneu) { 55533e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 55543e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5555312be037SStefano Zampini } 555630368db7SStefano Zampini if (nisvert) { 555730368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 555830368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 555930368db7SStefano Zampini } 5560f9eb5b7dSStefano Zampini 5561f9eb5b7dSStefano Zampini /* get some info after set from options */ 5562f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5563f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 55644f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 55656e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5566f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5567f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5568f9eb5b7dSStefano Zampini } 556939f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 55704f3a063dSStefano Zampini if (isredundant) { 55714f3a063dSStefano Zampini KSP inner_ksp; 55724f3a063dSStefano Zampini PC inner_pc; 55734f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 55744f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 55754f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 55764f3a063dSStefano Zampini } 5577f9eb5b7dSStefano Zampini 5578f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5579fa7f1dd8SStefano Zampini if (coarse_reuse) { 558081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5581fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 55826e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5583fa7f1dd8SStefano Zampini } else { 55846e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5585fa7f1dd8SStefano Zampini } 5586c8587f34SStefano Zampini if (isbddc || isnn) { 5587d4d8cf7bSStefano Zampini if (isbddc) { /* currently there's no API for this */ 5588d4d8cf7bSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc_temp->data; 5589d4d8cf7bSStefano Zampini pcbddc->detect_disconnected = PETSC_TRUE; 5590d4d8cf7bSStefano Zampini } 559122bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 559270cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5593b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 559422b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 55956e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 55976e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 55986e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 559922b6e8a2SStefano Zampini } 560070cf5478SStefano Zampini } 560153a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 560270cf5478SStefano Zampini } else { 560322bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 560422bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 560522bc73bbSStefano Zampini } 560622bc73bbSStefano Zampini } else { 56072e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5608c8587f34SStefano Zampini } 5609c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5610c8587f34SStefano Zampini 56113301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 56125a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 56133301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 56143301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 56153301b35fSStefano Zampini } 56163301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 56173301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 56183301b35fSStefano Zampini } 56193301b35fSStefano Zampini if (pc->pmat->spd_set) { 56203301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 56213301b35fSStefano Zampini } 56226e683305SStefano Zampini /* set operators */ 56235f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 56246e683305SStefano Zampini if (pcbddc->dbg_flag) { 56256e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56266e683305SStefano Zampini } 56276e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 56286e683305SStefano Zampini coarse_mat = 0; 56296e683305SStefano Zampini } 56306e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5631b1ecc7b1SStefano Zampini #if 0 5632b9b85e73SStefano Zampini { 5633b9b85e73SStefano Zampini PetscViewer viewer; 5634b9b85e73SStefano Zampini char filename[256]; 5635b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5636b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5637b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5638b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5639b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5640b9b85e73SStefano Zampini } 5641b9b85e73SStefano Zampini #endif 5642c8587f34SStefano Zampini 5643c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5644298c0119SStefano Zampini #if 0 5645c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5646c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 564798a51de6SStefano Zampini } 5648298c0119SStefano Zampini #endif 5649b0f5fe93SStefano Zampini /* hack */ 565098a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 565198a51de6SStefano Zampini Vec crhs,csol; 565204708bb6SStefano Zampini 5653f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5654f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5655f347579bSStefano Zampini if (!csol) { 56562a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5657f9eb5b7dSStefano Zampini } 5658f347579bSStefano Zampini if (!crhs) { 56592a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5660f347579bSStefano Zampini } 5661b0f5fe93SStefano Zampini } 5662b0f5fe93SStefano Zampini 5663b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5664b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5665b0f5fe93SStefano Zampini 5666b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 56674f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 56684f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 56694f1b2e48SStefano Zampini } 5670b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5671b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5672b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5673b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5674b0f5fe93SStefano Zampini if (coarse_mat) { 5675b0f5fe93SStefano Zampini Vec nullv; 5676b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5677b0f5fe93SStefano Zampini PetscInt nl; 5678b0f5fe93SStefano Zampini 5679b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5680b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5681b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5682b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5683b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5684b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5685b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5686b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5687b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5688b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5689b0f5fe93SStefano Zampini } 5690b0f5fe93SStefano Zampini } 5691b0f5fe93SStefano Zampini 5692b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5693b0f5fe93SStefano Zampini PetscBool ispreonly; 5694b0f5fe93SStefano Zampini 5695b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5696b0f5fe93SStefano Zampini PetscBool isnull; 5697b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5698d4d8cf7bSStefano Zampini if (0) { 569930368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5700b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5701b0f5fe93SStefano Zampini } else { 5702b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5703b0f5fe93SStefano Zampini } 5704b0f5fe93SStefano Zampini } else { 5705b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5706b0f5fe93SStefano Zampini } 5707b0f5fe93SStefano Zampini } 5708b0f5fe93SStefano Zampini /* setup coarse ksp */ 5709b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5710cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5711cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 57126e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5713c8587f34SStefano Zampini KSP check_ksp; 57142b510759SStefano Zampini KSPType check_ksp_type; 5715c8587f34SStefano Zampini PC check_pc; 57166e683305SStefano Zampini Vec check_vec,coarse_vec; 57176a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 57182b510759SStefano Zampini PetscInt its; 57196e683305SStefano Zampini PetscBool compute_eigs; 57206e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 57216e683305SStefano Zampini PetscInt neigs; 57228e185a42SStefano Zampini const char *prefix; 5723c8587f34SStefano Zampini 57242b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 57256e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5726422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 572723ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5728f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 57292b510759SStefano Zampini if (ispreonly) { 57302b510759SStefano Zampini check_ksp_type = KSPPREONLY; 57316e683305SStefano Zampini compute_eigs = PETSC_FALSE; 57322b510759SStefano Zampini } else { 5733cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 57346e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5735c8587f34SStefano Zampini } 5736c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 57376e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 57386e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 57396e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5740a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5741a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5742a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5743a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5744c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5745c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5746c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5747c8587f34SStefano Zampini /* create random vec */ 57486e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 57496e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5750c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5751c8587f34SStefano Zampini if (CoarseNullSpace) { 5752c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5753c8587f34SStefano Zampini } 57546e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5755c8587f34SStefano Zampini /* solve coarse problem */ 57566e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5757c8587f34SStefano Zampini if (CoarseNullSpace) { 57586e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5759c8587f34SStefano Zampini } 5760cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 57616e683305SStefano Zampini if (compute_eigs) { 5762854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5763854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 57646e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 57656e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 57666e683305SStefano Zampini lambda_min = eigs_r[0]; 57676e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 57686e683305SStefano Zampini if (lambda_max>lambda_min) { 5769cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5770cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5771cbcc2c2aSStefano Zampini } 5772c8587f34SStefano Zampini } 5773c8587f34SStefano Zampini } 5774cbcc2c2aSStefano Zampini 5775c8587f34SStefano Zampini /* check coarse problem residual error */ 57766e683305SStefano Zampini if (pcbddc->dbg_flag) { 57776e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 57786e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 57796e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5780c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 57816e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 57826e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5783c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5784779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 57856e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 57866e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 57876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 57886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5789b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5790b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5791b0f5fe93SStefano Zampini } 57926e683305SStefano Zampini if (compute_eigs) { 57936e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5794deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5795c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 57966e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 57976e683305SStefano 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); 57986e683305SStefano Zampini for (i=0;i<neigs;i++) { 57996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5800c8587f34SStefano Zampini } 58016e683305SStefano Zampini } 58026e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 58036e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58046e683305SStefano Zampini } 5805c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 58066e683305SStefano Zampini if (compute_eigs) { 58076e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 58086e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5809c8587f34SStefano Zampini } 58106e683305SStefano Zampini } 58116e683305SStefano Zampini } 5812cbcc2c2aSStefano Zampini /* print additional info */ 5813cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 58146e683305SStefano Zampini /* waits until all processes reaches this point */ 58156e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5816cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5817cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5818cbcc2c2aSStefano Zampini } 5819cbcc2c2aSStefano Zampini 58202b510759SStefano Zampini /* free memory */ 5821c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5822fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5823c8587f34SStefano Zampini PetscFunctionReturn(0); 5824c8587f34SStefano Zampini } 5825674ae819SStefano Zampini 5826f34684f1SStefano Zampini #undef __FUNCT__ 5827f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5828f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5829f34684f1SStefano Zampini { 5830f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5831f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5832f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5833dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5834dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 583573be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5836dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5837f34684f1SStefano Zampini PetscErrorCode ierr; 5838f34684f1SStefano Zampini 5839f34684f1SStefano Zampini PetscFunctionBegin; 5840f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 58410e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 58420e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5843727cdba6SStefano Zampini } 5844dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 58453bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5846dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5847dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5848dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5849dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5850dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5851dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 58520e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 58530e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 58540e6343abSStefano Zampini } 5855dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5856dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5857dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5858dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5859dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5860f34684f1SStefano Zampini 5861f34684f1SStefano Zampini /* check numbering */ 5862f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5863019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5864dc456d91SStefano Zampini PetscInt i; 5865b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5866f34684f1SStefano Zampini 5867f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5868f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5869f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 58701575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5871019a44ceSStefano Zampini /* counter */ 5872019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5873019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5874019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5875019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5876019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5877019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5878f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5879f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5880727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5881f34684f1SStefano Zampini } 5882f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5883f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5884f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5885e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5886e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5887e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5888e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5889f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5890019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5891f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5892019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 589375c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 589475c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5895b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5896019a44ceSStefano 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); 5897f34684f1SStefano Zampini } 5898f34684f1SStefano Zampini } 5899019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5900b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5901f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5902f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5903f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5904f34684f1SStefano Zampini } 5905f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5906f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5907e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5908e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5909f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5910f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5911b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5912ca8b9ea9SStefano Zampini PetscInt *gidxs; 5913ca8b9ea9SStefano Zampini 5914ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 59153bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5916f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5917f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5918f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5919f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 59204bc2dc4bSStefano 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); 5921f34684f1SStefano Zampini } 5922f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5923ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5924f34684f1SStefano Zampini } 5925f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59261575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5927302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5928f34684f1SStefano Zampini } 59298bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5930f34684f1SStefano Zampini /* get back data */ 5931f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5932f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5933674ae819SStefano Zampini PetscFunctionReturn(0); 5934674ae819SStefano Zampini } 5935674ae819SStefano Zampini 5936e456f2a8SStefano Zampini #undef __FUNCT__ 5937e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5938a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5939e456f2a8SStefano Zampini { 5940e456f2a8SStefano Zampini IS localis_t; 5941a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5942e456f2a8SStefano Zampini PetscScalar *vals; 5943e456f2a8SStefano Zampini PetscErrorCode ierr; 5944e456f2a8SStefano Zampini 5945e456f2a8SStefano Zampini PetscFunctionBegin; 5946a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5947e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5948854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5949e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5950e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5951a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5952a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 59531035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5954a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 59551035eff8SStefano Zampini } 5956a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5957e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5958e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5959a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5960a7dc3881SStefano Zampini /* now compute set in local ordering */ 5961a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5962a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5963a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5964a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5965a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5966ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5967e456f2a8SStefano Zampini lsize++; 5968e456f2a8SStefano Zampini } 5969e456f2a8SStefano Zampini } 5970854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 5971a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5972ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5973e456f2a8SStefano Zampini idxs[lsize++] = i; 5974e456f2a8SStefano Zampini } 5975e456f2a8SStefano Zampini } 5976a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5977a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 5978e456f2a8SStefano Zampini *localis = localis_t; 5979e456f2a8SStefano Zampini PetscFunctionReturn(0); 5980e456f2a8SStefano Zampini } 5981906d46d4SStefano Zampini 5982906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 5983906d46d4SStefano Zampini #undef __FUNCT__ 5984906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 5985906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 5986906d46d4SStefano Zampini { 5987906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 5988906d46d4SStefano Zampini PetscErrorCode ierr; 5989906d46d4SStefano Zampini 5990906d46d4SStefano Zampini PetscFunctionBegin; 5991906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 5992906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 5993906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 5994906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 5995906d46d4SStefano Zampini PetscFunctionReturn(0); 5996906d46d4SStefano Zampini } 5997906d46d4SStefano Zampini 5998906d46d4SStefano Zampini #undef __FUNCT__ 5999906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6000906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6001906d46d4SStefano Zampini { 6002906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6003906d46d4SStefano Zampini PetscErrorCode ierr; 6004906d46d4SStefano Zampini 6005906d46d4SStefano Zampini PetscFunctionBegin; 6006906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6007906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6008906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6009906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6010906d46d4SStefano Zampini PetscFunctionReturn(0); 6011906d46d4SStefano Zampini } 6012b96c3477SStefano Zampini 6013b96c3477SStefano Zampini #undef __FUNCT__ 6014b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 601508122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6016b96c3477SStefano Zampini { 6017a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6018b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6019b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6020a64f4aa4SStefano Zampini Mat S_j; 6021b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6022b96c3477SStefano Zampini PetscBool free_used_adj; 6023b96c3477SStefano Zampini PetscErrorCode ierr; 6024b96c3477SStefano Zampini 6025b96c3477SStefano Zampini PetscFunctionBegin; 6026b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6027b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 602808122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6029b96c3477SStefano Zampini used_xadj = NULL; 6030b96c3477SStefano Zampini used_adjncy = NULL; 6031b96c3477SStefano Zampini } else { 603208122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 603308122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 603408122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 603508122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6036b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6037b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6038b96c3477SStefano Zampini } else { 60392fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6040b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6041b96c3477SStefano Zampini PetscInt nvtxs; 6042b96c3477SStefano Zampini 60432fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 60442fffb893SStefano Zampini if (flg_row) { 6045b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6046b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6047b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6048b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 60492fffb893SStefano Zampini } else { 60502fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 60512fffb893SStefano Zampini used_xadj = NULL; 60522fffb893SStefano Zampini used_adjncy = NULL; 60532fffb893SStefano Zampini } 60542fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6055b96c3477SStefano Zampini } 6056b96c3477SStefano Zampini } 6057d5574798SStefano Zampini 6058d5574798SStefano Zampini /* setup sub_schurs data */ 6059a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6060df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6061df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6062a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6063ca92afb2SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL);CHKERRQ(ierr); 6064a64f4aa4SStefano Zampini } else { 60656816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 606604708bb6SStefano Zampini PetscBool isseqaij; 6067a3df083aSStefano Zampini PetscInt benign_n; 6068a3df083aSStefano Zampini 60695feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 60705feab87aSStefano Zampini PetscInt n_vertices; 60715feab87aSStefano Zampini 60725feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 60732034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 60745feab87aSStefano Zampini } 607504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 607604708bb6SStefano Zampini if (!isseqaij) { 607704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 607804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 607904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 608004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 608104708bb6SStefano Zampini } else { 608204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 608304708bb6SStefano Zampini } 608404708bb6SStefano Zampini } 6085a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6086a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6087ca92afb2SStefano Zampini } else { 6088a3df083aSStefano Zampini benign_n = 0; 6089ca92afb2SStefano Zampini } 6090a3df083aSStefano 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,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 6091ca92afb2SStefano Zampini } 6092d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6093b96c3477SStefano Zampini 6094b96c3477SStefano Zampini /* free adjacency */ 6095b96c3477SStefano Zampini if (free_used_adj) { 6096b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6097b96c3477SStefano Zampini } 6098b96c3477SStefano Zampini PetscFunctionReturn(0); 6099b96c3477SStefano Zampini } 6100b96c3477SStefano Zampini 6101b96c3477SStefano Zampini #undef __FUNCT__ 6102b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 610308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6104b96c3477SStefano Zampini { 6105b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6106b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6107b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6108b96c3477SStefano Zampini PCBDDCGraph graph; 6109b96c3477SStefano Zampini PetscErrorCode ierr; 6110b96c3477SStefano Zampini 6111b96c3477SStefano Zampini PetscFunctionBegin; 6112b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 611308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 61143301b35fSStefano Zampini IS verticesIS,verticescomm; 61153301b35fSStefano Zampini PetscInt vsize,*idxs; 6116b96c3477SStefano Zampini 6117b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 61183301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 61193301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61203301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 61213301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61223301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6123b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 61247fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 61253301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 61263301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6127b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6128b96c3477SStefano Zampini /* 6129b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6130b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6131b96c3477SStefano Zampini } 6132b96c3477SStefano Zampini */ 6133b96c3477SStefano Zampini } else { 6134b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6135b96c3477SStefano Zampini } 6136b96c3477SStefano Zampini 6137b96c3477SStefano Zampini /* sub_schurs init */ 6138a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6139a64f4aa4SStefano Zampini 6140b96c3477SStefano Zampini /* free graph struct */ 614108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6142b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6143b96c3477SStefano Zampini } 6144b96c3477SStefano Zampini PetscFunctionReturn(0); 6145b96c3477SStefano Zampini } 6146fa34dd3eSStefano Zampini 6147fa34dd3eSStefano Zampini #undef __FUNCT__ 6148fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6149fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6150fa34dd3eSStefano Zampini { 6151fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6152fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6153fa34dd3eSStefano Zampini PetscErrorCode ierr; 6154fa34dd3eSStefano Zampini 6155fa34dd3eSStefano Zampini PetscFunctionBegin; 6156fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6157fa34dd3eSStefano Zampini IS zerodiag = NULL; 61584f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6159fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 61604f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 616175c01103SStefano Zampini PetscReal norm; 6162fa34dd3eSStefano Zampini PetscInt i; 6163fa34dd3eSStefano Zampini 6164fa34dd3eSStefano Zampini /* B0 and B0_B */ 6165fa34dd3eSStefano Zampini if (zerodiag) { 6166fa34dd3eSStefano Zampini IS dummy; 6167fa34dd3eSStefano Zampini 61684f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 61694f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6170fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6171fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6172fa34dd3eSStefano Zampini } 6173fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6174fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6175fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6176fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6177fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6178fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6179fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6180fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6181fa34dd3eSStefano Zampini /* S_j */ 6182fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6183fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6184fa34dd3eSStefano Zampini 6185fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6186fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6187fa34dd3eSStefano Zampini /* continuous in primal space */ 6188fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6189fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6190fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6191fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 61924f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 61934f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6194fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6195fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6196fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6197fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6198fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6199fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6200fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6201fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6202fa34dd3eSStefano Zampini 6203fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6204fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6205fa34dd3eSStefano Zampini /* local with Schur */ 6206fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6207fa34dd3eSStefano Zampini if (zerodiag) { 6208fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62094f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6210fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6211fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6212fa34dd3eSStefano Zampini } 6213fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6214fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6215fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6216fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6217fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6218fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6219fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6220fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6221fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6222fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6223fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6224fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6225fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6226fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6227fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6228fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6229fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6230fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6231fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6232fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6233fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6234fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6235fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6236fa34dd3eSStefano Zampini if (zerodiag) { 6237fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6238fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62394f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6240fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6241fa34dd3eSStefano Zampini } 6242fa34dd3eSStefano Zampini /* BDDC */ 6243fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6244fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6245fa34dd3eSStefano Zampini 6246fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6247fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6248fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6249fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 62504f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 62514f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6252fa34dd3eSStefano Zampini } 62534f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6254fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6255fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6256fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6257fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6258fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6259fa34dd3eSStefano Zampini } 6260fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6261fa34dd3eSStefano Zampini } 6262