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; 2894f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 2904f1b2e48SStefano Zampini if (!isseqaij && filter) { 2911cf9b237SStefano Zampini PetscBool isseqdense; 2921cf9b237SStefano Zampini 2931cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 2941cf9b237SStefano Zampini if (!isseqdense) { 2954f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 2961cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 2971cf9b237SStefano Zampini PetscScalar *array; 2981cf9b237SStefano Zampini PetscReal chop=1.e-6; 2991cf9b237SStefano Zampini 3001cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3011cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3021cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3031cf9b237SStefano Zampini for (i=0;i<n;i++) { 3041cf9b237SStefano Zampini PetscInt j; 3051cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3061cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3071cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3081cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3091cf9b237SStefano Zampini } 3101cf9b237SStefano Zampini } 3111cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3121cf9b237SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); 3131cf9b237SStefano Zampini } 3144f1b2e48SStefano Zampini } else { 3154f1b2e48SStefano Zampini B = A; 3164f1b2e48SStefano Zampini } 3174f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3184f1b2e48SStefano Zampini 3194f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3204f1b2e48SStefano Zampini if (filter) { 3214f1b2e48SStefano Zampini PetscScalar *data; 3224f1b2e48SStefano Zampini PetscInt j,cum; 3234f1b2e48SStefano Zampini 3244f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3254f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3264f1b2e48SStefano Zampini cum = 0; 3274f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3284f1b2e48SStefano Zampini PetscInt t; 3294f1b2e48SStefano Zampini 3304f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3314f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3324f1b2e48SStefano Zampini continue; 3334f1b2e48SStefano Zampini } 3344f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3354f1b2e48SStefano Zampini } 3364f1b2e48SStefano Zampini t = xadj_filtered[i]; 3374f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3384f1b2e48SStefano Zampini cum += t; 3394f1b2e48SStefano Zampini } 3404f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3414f1b2e48SStefano Zampini } else { 3424f1b2e48SStefano Zampini xadj_filtered = NULL; 3434f1b2e48SStefano Zampini adjncy_filtered = NULL; 3444f1b2e48SStefano Zampini } 3454f1b2e48SStefano Zampini 3464f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3474f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3484f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3494f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3504f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3514f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3524f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3534f1b2e48SStefano Zampini if (xadj_filtered) { 3544f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3554f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3564f1b2e48SStefano Zampini } else { 3574f1b2e48SStefano Zampini graph->xadj = xadj; 3584f1b2e48SStefano Zampini graph->adjncy = adjncy; 3594f1b2e48SStefano Zampini } 3604f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3614f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3624f1b2e48SStefano Zampini /* partial clean up */ 3634f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3644f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3651cf9b237SStefano Zampini if (A != B) { 3664f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3674f1b2e48SStefano Zampini } 3684f1b2e48SStefano Zampini 3694f1b2e48SStefano Zampini /* get back data */ 3701cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3711cf9b237SStefano Zampini if (cc) { 3724f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3734f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3744f1b2e48SStefano 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); 3754f1b2e48SStefano Zampini } 3764f1b2e48SStefano Zampini *cc = cc_n; 3771cf9b237SStefano Zampini } 3784f1b2e48SStefano Zampini /* clean up graph */ 3794f1b2e48SStefano Zampini graph->xadj = 0; 3804f1b2e48SStefano Zampini graph->adjncy = 0; 3814f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 3824f1b2e48SStefano Zampini PetscFunctionReturn(0); 3834f1b2e48SStefano Zampini } 3844f1b2e48SStefano Zampini 3854f1b2e48SStefano Zampini #undef __FUNCT__ 3865408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 3875408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 3885408967cSStefano Zampini { 3895408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3905408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 391dee84bffSStefano Zampini IS dirIS = NULL; 3924f1b2e48SStefano Zampini PetscInt i; 3935408967cSStefano Zampini PetscErrorCode ierr; 3945408967cSStefano Zampini 3955408967cSStefano Zampini PetscFunctionBegin; 396dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 3975408967cSStefano Zampini if (zerodiag) { 3985408967cSStefano Zampini Mat A; 3995408967cSStefano Zampini Vec vec3_N; 4005408967cSStefano Zampini PetscScalar *vals; 4015408967cSStefano Zampini const PetscInt *idxs; 402*d12d3064SStefano Zampini PetscInt nz,*count; 4035408967cSStefano Zampini 4045408967cSStefano Zampini /* p0 */ 4055408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4065408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4075408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4085408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4094f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4105408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4115408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4125408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4135408967cSStefano Zampini /* v_I */ 4145408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4155408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4165408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4175408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4185408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4195408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4205408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4215408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4225408967cSStefano Zampini if (dirIS) { 4235408967cSStefano Zampini PetscInt n; 4245408967cSStefano Zampini 4255408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4265408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4275408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4285408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4295408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4305408967cSStefano Zampini } 4315408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4325408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4335408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4345408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4355408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4365408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4375408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 4384f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 4395408967cSStefano 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])); 4405408967cSStefano Zampini } 4415408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4425408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 443*d12d3064SStefano Zampini 444*d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 445*d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 446*d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 447*d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 448*d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 449*d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 450*d12d3064SStefano Zampini if (count[idxs[i]]) { 451*d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 452*d12d3064SStefano Zampini } 453*d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 454*d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4555408967cSStefano Zampini } 456dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4575408967cSStefano Zampini 4585408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4595408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4604f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4615408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4624f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4635408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4644f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4654f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4664f1b2e48SStefano 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); 4674f1b2e48SStefano Zampini } 4685408967cSStefano Zampini } 4695408967cSStefano Zampini PetscFunctionReturn(0); 4705408967cSStefano Zampini } 4715408967cSStefano Zampini 4725408967cSStefano Zampini #undef __FUNCT__ 473339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 474339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 475339f8db1SStefano Zampini { 476339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4774f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 478b0f5fe93SStefano Zampini PetscInt nz,n; 4794f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 480339f8db1SStefano Zampini PetscErrorCode ierr; 481339f8db1SStefano Zampini 482339f8db1SStefano Zampini PetscFunctionBegin; 4839f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4849f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 485339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 486a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 487a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 488a3df083aSStefano Zampini } 489a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 490a3df083aSStefano Zampini pcbddc->benign_n = 0; 4914f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4924f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 4934f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 4944f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 4954f1b2e48SStefano Zampini since the local Schur complements are SPD 4964f1b2e48SStefano Zampini */ 4974f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 4984f1b2e48SStefano Zampini have_null = PETSC_TRUE; 49940fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5004f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5014f1b2e48SStefano Zampini 5024f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5034f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5044f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5054f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 506ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 50740fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 50840fa8d13SStefano Zampini if (!sorted) { 50940fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 51040fa8d13SStefano Zampini } 51140fa8d13SStefano Zampini } else { 51240fa8d13SStefano Zampini pressures = NULL; 51340fa8d13SStefano Zampini } 51497d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 51597d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 51697d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 517339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 518339f8db1SStefano Zampini if (!sorted) { 519339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 520339f8db1SStefano Zampini } 521339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5224f1b2e48SStefano Zampini if (!nz) { 5234f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5244f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 52540fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 52640fa8d13SStefano Zampini } 5274f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5284f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5294f1b2e48SStefano Zampini zerodiag_subs = NULL; 5304f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5314f1b2e48SStefano Zampini if (has_null_pressures) { 5324f1b2e48SStefano Zampini IS *subs; 5334f1b2e48SStefano Zampini PetscInt nsubs,i; 5344f1b2e48SStefano Zampini 5354f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5364f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5374f1b2e48SStefano Zampini if (nsubs > 1) { 5384f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5394f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5404f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5414f1b2e48SStefano Zampini IS t_zerodiag_subs; 5424f1b2e48SStefano Zampini PetscInt nl; 5434f1b2e48SStefano Zampini 5444f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5454f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5464f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5474f1b2e48SStefano Zampini if (nl) { 5484f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5494f1b2e48SStefano Zampini 5504f1b2e48SStefano Zampini if (pressures) { 5514f1b2e48SStefano Zampini IS t_pressure_subs; 5524f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5534f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5544f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5554f1b2e48SStefano Zampini } 5564f1b2e48SStefano Zampini if (valid) { 5574f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5584f1b2e48SStefano Zampini pcbddc->benign_n++; 5594f1b2e48SStefano Zampini } else { 5604f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5614f1b2e48SStefano Zampini } 5624f1b2e48SStefano Zampini } 5634f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5644f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5654f1b2e48SStefano Zampini } 5664f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5674f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5684f1b2e48SStefano Zampini if (pressures) { 5694f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5704f1b2e48SStefano Zampini } 5714f1b2e48SStefano Zampini if (valid) { 5724f1b2e48SStefano Zampini pcbddc->benign_n = 1; 573ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5744f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5754f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5764f1b2e48SStefano Zampini } 5774f1b2e48SStefano Zampini } 5784f1b2e48SStefano Zampini } 5794f1b2e48SStefano Zampini 5804f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 5814f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5824f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5834f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5844f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5854f1b2e48SStefano Zampini } 5864f1b2e48SStefano Zampini 5874f1b2e48SStefano Zampini /* final check for null pressures */ 5884f1b2e48SStefano Zampini if (zerodiag && pressures) { 5894f1b2e48SStefano Zampini PetscInt nz,np; 5904f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5914f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 5924f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 5934f1b2e48SStefano Zampini } 5944f1b2e48SStefano Zampini 5954f1b2e48SStefano Zampini if (recompute_zerodiag) { 5964f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5974f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 5984f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 5994f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6004f1b2e48SStefano Zampini } else { 6014f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6024f1b2e48SStefano Zampini 6034f1b2e48SStefano Zampini nzn = 0; 6044f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6054f1b2e48SStefano Zampini PetscInt ns; 6064f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6074f1b2e48SStefano Zampini nzn += ns; 6084f1b2e48SStefano Zampini } 6094f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6104f1b2e48SStefano Zampini nzn = 0; 6114f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6124f1b2e48SStefano Zampini PetscInt ns,*idxs; 6134f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6144f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6154f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6164f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6174f1b2e48SStefano Zampini nzn += ns; 6184f1b2e48SStefano Zampini } 6194f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6204f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini } 6224f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6234f1b2e48SStefano Zampini } 6244f1b2e48SStefano Zampini 6254f1b2e48SStefano Zampini if (has_null_pressures) { 6264f1b2e48SStefano Zampini IS zerodiagc; 6274f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 6284f1b2e48SStefano Zampini PetscInt i,s,*nnz; 6294f1b2e48SStefano Zampini 6304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 631339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 632339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 633339f8db1SStefano Zampini /* local change of basis for pressures */ 634339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 63597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 636339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 637339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 638339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 6394f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 6404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6414f1b2e48SStefano Zampini PetscInt nzs,j; 6424f1b2e48SStefano Zampini 6434f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 6444f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6454f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 6464f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 6474f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6484f1b2e48SStefano Zampini } 649339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 650339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 651339f8db1SStefano Zampini /* set identity on velocities */ 652339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 653339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 654339f8db1SStefano Zampini } 6554f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 6564f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 6579f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 6584f1b2e48SStefano 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); 659339f8db1SStefano Zampini /* set change on pressures */ 6604f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 6614f1b2e48SStefano Zampini PetscScalar *array; 6624f1b2e48SStefano Zampini PetscInt nzs; 6634f1b2e48SStefano Zampini 6644f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 6654f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 6664f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 667339f8db1SStefano Zampini PetscScalar vals[2]; 668339f8db1SStefano Zampini PetscInt cols[2]; 669339f8db1SStefano Zampini 670339f8db1SStefano Zampini cols[0] = idxs[i]; 6714f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 672339f8db1SStefano Zampini vals[0] = 1.; 673b0f5fe93SStefano Zampini vals[1] = 1.; 6744f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 675339f8db1SStefano Zampini } 6764f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 6774f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 6784f1b2e48SStefano Zampini array[nzs-1] = 1.; 6794f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 6804f1b2e48SStefano Zampini /* store local idxs for p0 */ 6814f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 6824f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 683339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 6844f1b2e48SStefano Zampini } 685339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 686339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 687a3df083aSStefano Zampini /* project if needed */ 688a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 68997d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 690339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 69197d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 69297d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 693a3df083aSStefano Zampini } 6944f1b2e48SStefano Zampini /* store global idxs for p0 */ 6954f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 696339f8db1SStefano Zampini } 697ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 6984f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 699b0f5fe93SStefano Zampini 700b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 701b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 702339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 703339f8db1SStefano Zampini PetscFunctionReturn(0); 704339f8db1SStefano Zampini } 705339f8db1SStefano Zampini 706339f8db1SStefano Zampini #undef __FUNCT__ 707015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 708015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 709efc2fbd9SStefano Zampini { 710efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 711efc2fbd9SStefano Zampini PetscErrorCode ierr; 712efc2fbd9SStefano Zampini 713efc2fbd9SStefano Zampini PetscFunctionBegin; 714efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 715efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 7164f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 717efc2fbd9SStefano Zampini } 718015636ebSStefano Zampini if (get) { /* use SF to get values */ 719efc2fbd9SStefano Zampini PetscScalar *array; 720efc2fbd9SStefano Zampini 721efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7224f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 7234f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 724efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7254f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 7264f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 727efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 728efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 729efc2fbd9SStefano Zampini } 730efc2fbd9SStefano Zampini PetscFunctionReturn(0); 731efc2fbd9SStefano Zampini } 732efc2fbd9SStefano Zampini 733efc2fbd9SStefano Zampini #undef __FUNCT__ 734c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 735c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 736c263805aSStefano Zampini { 737c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 738c263805aSStefano Zampini PetscErrorCode ierr; 739c263805aSStefano Zampini 740c263805aSStefano Zampini PetscFunctionBegin; 741c263805aSStefano Zampini /* TODO: add error checking 742c263805aSStefano Zampini - avoid nested pop (or push) calls. 743c263805aSStefano Zampini - cannot push before pop. 7441c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 745c263805aSStefano Zampini */ 7464f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 747efc2fbd9SStefano Zampini PetscFunctionReturn(0); 748efc2fbd9SStefano Zampini } 749c263805aSStefano Zampini if (pop) { 750a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7514f1b2e48SStefano Zampini IS is_p0; 7524f1b2e48SStefano Zampini MatReuse reuse; 753c263805aSStefano Zampini 754c263805aSStefano Zampini /* extract B_0 */ 7554f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 7564f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 7574f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 7584f1b2e48SStefano Zampini } 7594f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 7604f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 761c263805aSStefano Zampini /* remove rows and cols from local problem */ 762c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 76397d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 7644f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 7654f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 766a3df083aSStefano Zampini } else { 767a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 768a3df083aSStefano Zampini PetscScalar *vals; 769a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 770a3df083aSStefano Zampini 771a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 772a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 773a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 774a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 775a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 776a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 777a3df083aSStefano Zampini /* this matrix is very sparse: the nnz pattern is not known unless we do 2 sweeps of the next loop. 778a3df083aSStefano Zampini Setting nnz=100 should be more than enough */ 779a3df083aSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,100,NULL);CHKERRQ(ierr); 780a3df083aSStefano Zampini } 781a3df083aSStefano Zampini 782a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 783a3df083aSStefano Zampini PetscScalar *array; 784a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 785a3df083aSStefano Zampini 786a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 787a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 788a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 789a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 790a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 791a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 792a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 793a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 794a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 795a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 796a3df083aSStefano Zampini cum = 0; 797a3df083aSStefano Zampini for (j=0;j<n;j++) { 798a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 799a3df083aSStefano Zampini vals[cum] = array[j]; 800a3df083aSStefano Zampini idxs_ins[cum] = j; 801a3df083aSStefano Zampini cum++; 802a3df083aSStefano Zampini } 803a3df083aSStefano Zampini } 804a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 805a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 806a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 807a3df083aSStefano Zampini } 808a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 809a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 810a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 811a3df083aSStefano Zampini } 812c263805aSStefano Zampini } else { /* push */ 813a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8144f1b2e48SStefano Zampini PetscInt i; 8154f1b2e48SStefano Zampini 8164f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8174f1b2e48SStefano Zampini PetscScalar *B0_vals; 8184f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 8194f1b2e48SStefano Zampini 8204f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8214f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8227b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8234f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 8244f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8254f1b2e48SStefano Zampini } 826c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 827c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 828a3df083aSStefano Zampini } else { 829a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 830a3df083aSStefano Zampini } 831c263805aSStefano Zampini } 832c263805aSStefano Zampini PetscFunctionReturn(0); 833c263805aSStefano Zampini } 834c263805aSStefano Zampini 835c263805aSStefano Zampini #undef __FUNCT__ 836b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 83708122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 838b1b3d7a2SStefano Zampini { 839b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 84008122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 84108122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 84208122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 84308122e43SStefano Zampini PetscScalar *work,lwork; 84408122e43SStefano Zampini PetscScalar *St,*S,*eigv; 84508122e43SStefano Zampini PetscScalar *Sarray,*Starray; 84608122e43SStefano Zampini PetscReal *eigs,thresh; 8471b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 848f6f667cfSStefano Zampini PetscBool allocated_S_St; 84908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 85008122e43SStefano Zampini PetscReal *rwork; 85108122e43SStefano Zampini #endif 852b1b3d7a2SStefano Zampini PetscErrorCode ierr; 853b1b3d7a2SStefano Zampini 854b1b3d7a2SStefano Zampini PetscFunctionBegin; 855df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 856df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 85708122e43SStefano Zampini } 85808122e43SStefano Zampini 85906a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 86006a4e24aSStefano 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); 86106a4e24aSStefano Zampini } 86206a4e24aSStefano Zampini 863fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 864fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 865fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 866fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 8671575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 868fd14bc51SStefano Zampini } 869fd14bc51SStefano Zampini 870e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 871e496cd5dSStefano 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); 872e496cd5dSStefano Zampini } 873e496cd5dSStefano Zampini 87408122e43SStefano Zampini /* max size of subsets */ 87508122e43SStefano Zampini mss = 0; 87608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 87708122e43SStefano Zampini PetscInt subset_size; 878862806e4SStefano Zampini 87908122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 88008122e43SStefano Zampini mss = PetscMax(mss,subset_size); 88108122e43SStefano Zampini } 88208122e43SStefano Zampini 88308122e43SStefano Zampini /* min/max and threshold */ 88408122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 885f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 88608122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 887f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 888f6f667cfSStefano Zampini if (nmin) { 889f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 890f6f667cfSStefano Zampini } 89108122e43SStefano Zampini 89208122e43SStefano Zampini /* allocate lapack workspace */ 89308122e43SStefano Zampini cum = cum2 = 0; 89408122e43SStefano Zampini maxneigs = 0; 89508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 89608122e43SStefano Zampini PetscInt n,subset_size; 897f6f667cfSStefano Zampini 89808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 89908122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9009162d606SStefano Zampini cum += subset_size; 9019162d606SStefano Zampini cum2 += subset_size*n; 90208122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 90308122e43SStefano Zampini } 90408122e43SStefano Zampini if (mss) { 9059ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 90608122e43SStefano Zampini PetscBLASInt B_itype = 1; 90708122e43SStefano Zampini PetscBLASInt B_N = mss; 9084c6709b3SStefano Zampini PetscReal zero = 0.0; 9094c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 91008122e43SStefano Zampini 91108122e43SStefano Zampini B_lwork = -1; 91208122e43SStefano Zampini S = NULL; 91308122e43SStefano Zampini St = NULL; 914a58a30b4SStefano Zampini eigs = NULL; 915a58a30b4SStefano Zampini eigv = NULL; 916a58a30b4SStefano Zampini B_iwork = NULL; 917a58a30b4SStefano Zampini B_ifail = NULL; 918d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 919d1710679SStefano Zampini rwork = NULL; 920d1710679SStefano Zampini #endif 9218bec7fa6SStefano Zampini thresh = 1.0; 92208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 92308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 92408122e43SStefano 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)); 92508122e43SStefano Zampini #else 92608122e43SStefano 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)); 92708122e43SStefano Zampini #endif 92808122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 92908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 93008122e43SStefano Zampini } else { 93108122e43SStefano Zampini /* TODO */ 93208122e43SStefano Zampini } 93308122e43SStefano Zampini } else { 93408122e43SStefano Zampini lwork = 0; 93508122e43SStefano Zampini } 93608122e43SStefano Zampini 93708122e43SStefano Zampini nv = 0; 938d62866d3SStefano 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) */ 939d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 94008122e43SStefano Zampini } 9414c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 942f6f667cfSStefano Zampini if (allocated_S_St) { 943f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 944f6f667cfSStefano Zampini } 945f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 94608122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 94708122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 94808122e43SStefano Zampini #endif 9499162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 9509162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 9519162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 95208122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 9539162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 95408122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 95508122e43SStefano Zampini 95608122e43SStefano Zampini maxneigs = 0; 95708122e43SStefano Zampini cum = cum2 = cumarray = 0; 9589162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 9599162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 960d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 96108122e43SStefano Zampini const PetscInt *idxs; 96208122e43SStefano Zampini 963d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 96408122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 96508122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 96608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 96708122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 9689162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 9699162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 97008122e43SStefano Zampini } 97108122e43SStefano Zampini cum2 = cum; 972d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 97308122e43SStefano Zampini } 97408122e43SStefano Zampini 97508122e43SStefano Zampini if (mss) { /* multilevel */ 97608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 97708122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 97808122e43SStefano Zampini } 97908122e43SStefano Zampini 980ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 98108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 98208122e43SStefano Zampini const PetscInt *idxs; 983f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 984862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 98508122e43SStefano Zampini PetscBLASInt B_N; 986aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 98708122e43SStefano Zampini 988862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 989ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 990f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 991f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 9929ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 993aff50787SStefano Zampini PetscInt j,k; 994aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 995aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 996aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 99708122e43SStefano Zampini } 99808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 999aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1000aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1001aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1002aff50787SStefano Zampini } 100308122e43SStefano Zampini } 100408122e43SStefano Zampini } else { 100508122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 100608122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 100708122e43SStefano Zampini } 10088bec7fa6SStefano Zampini } else { 1009f6f667cfSStefano Zampini S = Sarray + cumarray; 1010f6f667cfSStefano Zampini St = Starray + cumarray; 10118bec7fa6SStefano Zampini } 101208122e43SStefano Zampini 1013aff50787SStefano Zampini /* see if we can save some work */ 1014aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1015aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1016aff50787SStefano Zampini } 1017aff50787SStefano Zampini 1018aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1019aff50787SStefano Zampini B_neigs = 0; 1020aff50787SStefano Zampini } else { 10219ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 102208122e43SStefano Zampini PetscBLASInt B_itype = 1; 1023f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 10244c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 10259552c7c7SStefano Zampini PetscInt nmin_s; 102608122e43SStefano Zampini 1027fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 10288bec7fa6SStefano 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]]); 1029fd14bc51SStefano Zampini } 1030d16cbb6bSStefano Zampini 103108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1032d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1033d16cbb6bSStefano Zampini 1034d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 103508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1036f6f667cfSStefano 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)); 103708122e43SStefano Zampini #else 1038f6f667cfSStefano 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)); 103908122e43SStefano Zampini #endif 1040d16cbb6bSStefano Zampini } else { 1041d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1042d16cbb6bSStefano Zampini B_IL = 1; 1043d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1044d16cbb6bSStefano 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)); 1045d16cbb6bSStefano Zampini #else 1046d16cbb6bSStefano 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)); 1047d16cbb6bSStefano Zampini #endif 1048d16cbb6bSStefano Zampini } 104908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 105008122e43SStefano Zampini if (B_ierr) { 105108122e43SStefano Zampini if (B_ierr < 0 ) { 105208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 105308122e43SStefano Zampini } else if (B_ierr <= B_N) { 105408122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 105508122e43SStefano Zampini } else { 10569552c7c7SStefano 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); 105708122e43SStefano Zampini } 105808122e43SStefano Zampini } 105908122e43SStefano Zampini 106008122e43SStefano Zampini if (B_neigs > nmax) { 1061fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1062fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1063fd14bc51SStefano Zampini } 1064f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 106508122e43SStefano Zampini B_neigs = nmax; 106608122e43SStefano Zampini } 106708122e43SStefano Zampini 10689552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 10699552c7c7SStefano Zampini if (B_neigs < nmin_s) { 107008122e43SStefano Zampini PetscBLASInt B_neigs2; 107108122e43SStefano Zampini 1072f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1073f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1074fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1075fd14bc51SStefano 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); 1076fd14bc51SStefano Zampini } 10779ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 107808122e43SStefano Zampini PetscInt j; 107908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 108008122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 108108122e43SStefano Zampini } 108208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 108308122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 108408122e43SStefano Zampini } 108508122e43SStefano Zampini } else { 108608122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 108708122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 108808122e43SStefano Zampini } 108908122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 109008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1091f6f667cfSStefano 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)); 109208122e43SStefano Zampini #else 1093f6f667cfSStefano 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)); 109408122e43SStefano Zampini #endif 109508122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 109608122e43SStefano Zampini B_neigs += B_neigs2; 109708122e43SStefano Zampini } 109808122e43SStefano Zampini if (B_ierr) { 109908122e43SStefano Zampini if (B_ierr < 0 ) { 110008122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 110108122e43SStefano Zampini } else if (B_ierr <= B_N) { 110208122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 110308122e43SStefano Zampini } else { 11049552c7c7SStefano 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); 110508122e43SStefano Zampini } 110608122e43SStefano Zampini } 1107fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1108ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 110908122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 111008122e43SStefano Zampini if (eigs[j] == 0.0) { 1111ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 111208122e43SStefano Zampini } else { 1113ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1114fd14bc51SStefano Zampini } 111508122e43SStefano Zampini } 111608122e43SStefano Zampini } 111708122e43SStefano Zampini } else { 111808122e43SStefano Zampini /* TODO */ 111908122e43SStefano Zampini } 1120aff50787SStefano Zampini } 11218bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 11228bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 11239162d606SStefano Zampini if (B_neigs) { 11249162d606SStefano 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); 1125fd14bc51SStefano Zampini 1126fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 11279552c7c7SStefano Zampini PetscInt ii; 11289552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1129ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11309552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1131ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1132ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1133ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1134ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1135ac47001eSStefano Zampini #else 1136ac47001eSStefano 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); 1137ac47001eSStefano Zampini #endif 11389552c7c7SStefano Zampini } 11399552c7c7SStefano Zampini } 1140fd14bc51SStefano Zampini } 114108122e43SStefano Zampini #if 0 11429162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 114308122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 114408122e43SStefano Zampini PetscScalar norm; 114508122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 11469162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 11479162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 114808122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 114908122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 115008122e43SStefano Zampini } else { 115108122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 115208122e43SStefano Zampini } 11539162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 1154b1b3d7a2SStefano Zampini } 1155b1b3d7a2SStefano Zampini #endif 11569162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 11579162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 11589162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 11599162d606SStefano Zampini cum++; 116008122e43SStefano Zampini } 116108122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 116208122e43SStefano Zampini /* shift for next computation */ 116308122e43SStefano Zampini cumarray += subset_size*subset_size; 116408122e43SStefano Zampini } 1165fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1166fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1167fd14bc51SStefano Zampini } 116808122e43SStefano Zampini 116908122e43SStefano Zampini if (mss) { 117008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 117108122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1172f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1173f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1174f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 117508122e43SStefano Zampini } 1176f6f667cfSStefano Zampini if (allocated_S_St) { 1177f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1178f6f667cfSStefano Zampini } 1179f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 118008122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 118108122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 118208122e43SStefano Zampini #endif 118308122e43SStefano Zampini if (pcbddc->dbg_flag) { 11841b968477SStefano Zampini PetscInt maxneigs_r; 118508122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 11869b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 118708122e43SStefano Zampini } 118808122e43SStefano Zampini PetscFunctionReturn(0); 118908122e43SStefano Zampini } 1190b1b3d7a2SStefano Zampini 1191674ae819SStefano Zampini #undef __FUNCT__ 1192c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1193c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1194c8587f34SStefano Zampini { 1195c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 11968629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1197c8587f34SStefano Zampini PetscErrorCode ierr; 1198c8587f34SStefano Zampini 1199c8587f34SStefano Zampini PetscFunctionBegin; 1200f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12015e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1202c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1203c8587f34SStefano Zampini 1204684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12050fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1206684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1207c8587f34SStefano Zampini 1208c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1209b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1210c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1211c8587f34SStefano Zampini } 1212c8587f34SStefano Zampini 12138629588bSStefano Zampini /* 12148629588bSStefano Zampini Setup local correction and local part of coarse basis. 12158629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12168629588bSStefano Zampini */ 121747f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12188629588bSStefano Zampini 12198629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12208629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12218629588bSStefano Zampini 12228629588bSStefano Zampini /* free */ 12238629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1224c8587f34SStefano Zampini PetscFunctionReturn(0); 1225c8587f34SStefano Zampini } 1226c8587f34SStefano Zampini 1227c8587f34SStefano Zampini #undef __FUNCT__ 1228674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1229674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1230674ae819SStefano Zampini { 1231674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1232674ae819SStefano Zampini PetscErrorCode ierr; 1233674ae819SStefano Zampini 1234674ae819SStefano Zampini PetscFunctionBegin; 1235674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1236674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 123730368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1238674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1239674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1240785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1241674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1242f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1243f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1244785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 124563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 124663602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1247674ae819SStefano Zampini PetscFunctionReturn(0); 1248674ae819SStefano Zampini } 1249674ae819SStefano Zampini 1250674ae819SStefano Zampini #undef __FUNCT__ 1251674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1252674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1253674ae819SStefano Zampini { 1254674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12554f1b2e48SStefano Zampini PetscInt i; 1256674ae819SStefano Zampini PetscErrorCode ierr; 1257674ae819SStefano Zampini 1258674ae819SStefano Zampini PetscFunctionBegin; 1259b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1260674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1261674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1262674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 12634f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 12644f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 12654f1b2e48SStefano Zampini } 12664f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1267b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1268674ae819SStefano Zampini PetscFunctionReturn(0); 1269674ae819SStefano Zampini } 1270674ae819SStefano Zampini 1271674ae819SStefano Zampini #undef __FUNCT__ 1272674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1273674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1274674ae819SStefano Zampini { 1275674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1276674ae819SStefano Zampini PetscErrorCode ierr; 1277674ae819SStefano Zampini 1278674ae819SStefano Zampini PetscFunctionBegin; 1279674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 128058da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1281ca92afb2SStefano Zampini PetscScalar *array; 128206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 128306656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 128458da7f69SStefano Zampini } 1285674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1286674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 128715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 128815aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1289674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1290674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1291674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 129206656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1293674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1294674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 12958ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1296674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1297674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1298674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1299f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1300f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1301f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1302f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1303727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13040e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1305f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 130670cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 13076e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 130881d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 13090369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 13108b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 13114f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13128b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1313ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1314ca92afb2SStefano Zampini PetscInt i; 1315ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1316ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1317ca92afb2SStefano Zampini } 1318ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1319ca92afb2SStefano Zampini } 13204f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1321674ae819SStefano Zampini PetscFunctionReturn(0); 1322674ae819SStefano Zampini } 1323674ae819SStefano Zampini 1324674ae819SStefano Zampini #undef __FUNCT__ 1325f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1326f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 13276bfb1811SStefano Zampini { 13286bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13296bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 13306bfb1811SStefano Zampini VecType impVecType; 13314f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 13326bfb1811SStefano Zampini PetscErrorCode ierr; 13336bfb1811SStefano Zampini 13346bfb1811SStefano Zampini PetscFunctionBegin; 1335f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1336019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1337f4ddd8eeSStefano Zampini } 1338e7b262bdSStefano Zampini /* get sizes */ 13394f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1340b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 13416bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1342e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1343e7b262bdSStefano Zampini /* R nodes */ 1344e7b262bdSStefano Zampini old_size = -1; 1345e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1346e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1347e7b262bdSStefano Zampini } 1348e7b262bdSStefano Zampini if (n_R != old_size) { 1349e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1350e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13516bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 13526bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 13536bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 13546bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1355e7b262bdSStefano Zampini } 1356e7b262bdSStefano Zampini /* local primal dofs */ 1357e7b262bdSStefano Zampini old_size = -1; 1358e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1359e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1360e7b262bdSStefano Zampini } 1361e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1362e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 136383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1364e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 13656bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1366e7b262bdSStefano Zampini } 1367e7b262bdSStefano Zampini /* local explicit constraints */ 1368e7b262bdSStefano Zampini old_size = -1; 1369e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1370e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1371e7b262bdSStefano Zampini } 1372e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1373e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 137483b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 137583b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 137683b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 137783b7ccabSStefano Zampini } 13786bfb1811SStefano Zampini PetscFunctionReturn(0); 13796bfb1811SStefano Zampini } 13806bfb1811SStefano Zampini 13816bfb1811SStefano Zampini #undef __FUNCT__ 138247f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 138347f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 138488ebb749SStefano Zampini { 138525084f0cSStefano Zampini PetscErrorCode ierr; 138625084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 138788ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 138888ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1389d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 139025084f0cSStefano Zampini /* submatrices of local problem */ 139180677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1392019a44ceSStefano Zampini /* submatrices of benign trick */ 1393d16cbb6bSStefano Zampini Mat B0_V = NULL; 139406656605SStefano Zampini /* submatrices of local coarse problem */ 139506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 139625084f0cSStefano Zampini /* working matrices */ 139706656605SStefano Zampini Mat C_CR; 139825084f0cSStefano Zampini /* additional working stuff */ 139906656605SStefano Zampini PC pc_R; 14004f1b2e48SStefano Zampini Mat F; 1401a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 140206656605SStefano Zampini 140325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 140406656605SStefano Zampini PetscScalar *work; 140506656605SStefano Zampini PetscInt *idx_V_B; 1406ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 140706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1408ffd830a3SStefano Zampini 140925084f0cSStefano Zampini /* some shortcuts to scalars */ 141006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 141188ebb749SStefano Zampini 141288ebb749SStefano Zampini PetscFunctionBegin; 1413ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1414ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1415ffd830a3SStefano Zampini } 1416ffd830a3SStefano Zampini 1417ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1418b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14194f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1420b371cd4fSStefano Zampini n_B = pcis->n_B; 1421b371cd4fSStefano Zampini n_D = pcis->n - n_B; 142288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 142388ebb749SStefano Zampini 142488ebb749SStefano Zampini /* vertices in boundary numbering */ 1425785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 14260e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 142788ebb749SStefano Zampini if (i != n_vertices) { 142822d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 142988ebb749SStefano Zampini } 143088ebb749SStefano Zampini 143106656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1432019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 143306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 143406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 143506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 143606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 143706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 143806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 143906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 144006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 144106656605SStefano Zampini 144206656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 144306656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 144406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 144506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 144606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1447ffd830a3SStefano Zampini lda_rhs = n_R; 1448a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 144906656605SStefano Zampini if (isLU || isILU || isCHOL) { 145006656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1451df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1452df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1453d62866d3SStefano Zampini MatFactorType type; 1454d62866d3SStefano Zampini 1455df4d28bfSStefano Zampini F = reuse_solver->F; 14566816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1457d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1458ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1459df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 146006656605SStefano Zampini } else { 146106656605SStefano Zampini F = NULL; 146206656605SStefano Zampini } 146306656605SStefano Zampini 1464ffd830a3SStefano Zampini /* allocate workspace */ 1465ffd830a3SStefano Zampini n = 0; 1466ffd830a3SStefano Zampini if (n_constraints) { 1467ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1468ffd830a3SStefano Zampini } 1469ffd830a3SStefano Zampini if (n_vertices) { 1470ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1471ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1472ffd830a3SStefano Zampini } 1473ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1474ffd830a3SStefano Zampini 147588ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 147688ebb749SStefano Zampini if (n_constraints) { 147706656605SStefano Zampini Mat M1,M2,M3; 147880677318SStefano Zampini Mat auxmat; 147906656605SStefano Zampini IS is_aux; 148080677318SStefano Zampini PetscScalar *array,*array2; 148106656605SStefano Zampini 1482f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 148380677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 148488ebb749SStefano Zampini 148525084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 148625084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 14878ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 148880677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 148988ebb749SStefano Zampini 149080677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 149180677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1492ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 149388ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 149406656605SStefano Zampini const PetscScalar *row_cmat_values; 149506656605SStefano Zampini const PetscInt *row_cmat_indices; 149606656605SStefano Zampini PetscInt size_of_constraint,j; 149788ebb749SStefano Zampini 149806656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 149906656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1500ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 150106656605SStefano Zampini } 150206656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 150306656605SStefano Zampini } 1504ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 150506656605SStefano Zampini if (F) { 150606656605SStefano Zampini Mat B; 150706656605SStefano Zampini 1508ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1509a3df083aSStefano Zampini if (need_benign_correction) { 1510a3df083aSStefano Zampini PetscScalar *marr; 1511df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1512a3df083aSStefano Zampini 1513a3df083aSStefano Zampini ierr = MatDenseGetArray(B,&marr);CHKERRQ(ierr); 1514a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1515a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1516df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1517a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1518a3df083aSStefano Zampini } 1519a3df083aSStefano Zampini ierr = MatDenseRestoreArray(B,&marr);CHKERRQ(ierr); 1520a3df083aSStefano Zampini } 152180677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1522a3df083aSStefano Zampini if (need_benign_correction) { 1523a3df083aSStefano Zampini PetscScalar *marr; 1524df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1525a3df083aSStefano Zampini 1526a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1527a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1528a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1529df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1530a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1531a3df083aSStefano Zampini } 1532a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1533a3df083aSStefano Zampini } 153406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 153506656605SStefano Zampini } else { 153680677318SStefano Zampini PetscScalar *marr; 153780677318SStefano Zampini 153880677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 153906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1540ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1541ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 154206656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 154306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 154406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 154506656605SStefano Zampini } 154680677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 154706656605SStefano Zampini } 154880677318SStefano Zampini if (!pcbddc->switch_static) { 154980677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 155080677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 155180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 155280677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1553ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 155480677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 155580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 155680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 155780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 155880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 155980677318SStefano Zampini } 156080677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 156180677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 156280677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 156380677318SStefano Zampini } else { 1564ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1565ffd830a3SStefano Zampini IS dummy; 1566ffd830a3SStefano Zampini 1567ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1568ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1569ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1570ffd830a3SStefano Zampini } else { 157180677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 157280677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1573ffd830a3SStefano Zampini } 157425084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 157580677318SStefano Zampini } 157680677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 157780677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 157880677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 157906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 158006656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 158180677318SStefano Zampini if (isCHOL) { 158280677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 158380677318SStefano Zampini } else { 158425084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 158580677318SStefano Zampini } 158680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 158706656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 158825084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 158925084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 159025084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 159180677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 159280677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 159380677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 159406656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 159506656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1596f4ddd8eeSStefano Zampini } 159788ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 15984f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1599d16cbb6bSStefano Zampini IS dummy; 1600d16cbb6bSStefano Zampini Mat B0_R; 1601d16cbb6bSStefano Zampini PetscReal norm; 1602d16cbb6bSStefano Zampini 16034f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16044f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1605d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1606d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1607d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1608d16cbb6bSStefano Zampini } 1609d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1610d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1611d16cbb6bSStefano Zampini } 1612d16cbb6bSStefano Zampini 161388ebb749SStefano Zampini if (n_vertices) { 161406656605SStefano Zampini IS is_aux; 16153a50541eSStefano Zampini 1616df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16176816873aSStefano Zampini IS tis; 16186816873aSStefano Zampini 16196816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16206816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16216816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16226816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16236816873aSStefano Zampini } else { 16243a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16256816873aSStefano Zampini } 16269577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16279577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 162804708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 16294f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1630019a44ceSStefano Zampini IS dummy; 1631019a44ceSStefano Zampini 16324f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16334f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1634019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1635019a44ceSStefano Zampini } 163625084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 163788ebb749SStefano Zampini } 163888ebb749SStefano Zampini 163988ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1640f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 164106656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 164206656605SStefano Zampini if (pcbddc->coarse_phi_D) { 164306656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 164406656605SStefano Zampini } 1645f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 164606656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 164706656605SStefano Zampini PetscScalar *marray; 164806656605SStefano Zampini 164906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 165006656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1651f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1652f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1653f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1654f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1655f4ddd8eeSStefano Zampini } 1656f4ddd8eeSStefano Zampini } 165706656605SStefano Zampini 1658f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 165906656605SStefano Zampini PetscScalar *marray; 166088ebb749SStefano Zampini 166106656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16628eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 166306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 166488ebb749SStefano Zampini } 16653301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 166606656605SStefano Zampini n *= 2; 166788ebb749SStefano Zampini } 166806656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 166906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 167006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16718eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 167206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 167306656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 167488ebb749SStefano Zampini } 16753301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 167606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 16778eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 167806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 167906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 168088ebb749SStefano Zampini } 168188ebb749SStefano Zampini } else { 1682c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1683c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 16841b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1685c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1686c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1687c0553b1fSStefano Zampini } 168888ebb749SStefano Zampini } 168906656605SStefano Zampini } 1690019a44ceSStefano Zampini 169106656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 16924f1b2e48SStefano Zampini p0_lidx_I = NULL; 16934f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1694d12edf2fSStefano Zampini const PetscInt *idxs; 1695d12edf2fSStefano Zampini 1696d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 16974f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 16984f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 16994f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17004f1b2e48SStefano Zampini } 1701d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1702d12edf2fSStefano Zampini } 1703d16cbb6bSStefano Zampini 170406656605SStefano Zampini /* vertices */ 170506656605SStefano Zampini if (n_vertices) { 170616f15bc4SStefano Zampini 1707ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 170804708bb6SStefano Zampini 170916f15bc4SStefano Zampini if (n_R) { 171014393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 171106656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 171216f15bc4SStefano Zampini PetscScalar *x,*y; 171304708bb6SStefano Zampini PetscBool isseqaij; 171406656605SStefano Zampini 171521eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 171614393ed6SStefano Zampini if (need_benign_correction) { 171714393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 171814393ed6SStefano Zampini IS is_p0; 171914393ed6SStefano Zampini PetscInt *idxs_p0,n; 172014393ed6SStefano Zampini 172114393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 172214393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 172314393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 172414393ed6SStefano Zampini if (n != pcbddc->benign_n) { 172514393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 172614393ed6SStefano Zampini } 172714393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 172814393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 172914393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 173014393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 173114393ed6SStefano Zampini } 173214393ed6SStefano Zampini 1733ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1734ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1735ffd830a3SStefano Zampini } else { 1736ca92afb2SStefano Zampini PetscScalar *av,*array; 1737ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1738ca92afb2SStefano Zampini PetscInt n; 1739ca92afb2SStefano Zampini PetscBool flg_row; 1740ffd830a3SStefano Zampini 1741ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1742ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1743ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1744ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1745ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1746ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1747ca92afb2SStefano Zampini PetscInt j; 1748ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1749ffd830a3SStefano Zampini } 1750ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1751ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1752ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1753ffd830a3SStefano Zampini } 1754ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 175514393ed6SStefano Zampini if (F) { 1756a3df083aSStefano Zampini if (need_benign_correction) { 1757df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1758a3df083aSStefano Zampini PetscScalar *marr; 1759a3df083aSStefano Zampini 1760a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 176114393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 176214393ed6SStefano Zampini 176314393ed6SStefano Zampini | 0 0 0 | (V) 176414393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 176514393ed6SStefano Zampini | 0 0 -1 | (p0) 176614393ed6SStefano Zampini 176714393ed6SStefano Zampini */ 1768df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 176914393ed6SStefano Zampini const PetscScalar *vals; 177014393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 177114393ed6SStefano Zampini PetscInt n,j,nz; 177214393ed6SStefano Zampini 1773df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1774df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 177514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 177614393ed6SStefano Zampini for (j=0;j<n;j++) { 177714393ed6SStefano Zampini PetscScalar val = vals[j]; 177814393ed6SStefano Zampini PetscInt k,col = idxs[j]; 177914393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 178014393ed6SStefano Zampini } 178114393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1782df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 178314393ed6SStefano Zampini } 178414393ed6SStefano Zampini /* need to correct the rhs */ 1785a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1786a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1787df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1788a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1789a3df083aSStefano Zampini } 1790a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1791a3df083aSStefano Zampini } 179206656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 179314393ed6SStefano Zampini /* need to correct the solution */ 1794a3df083aSStefano Zampini if (need_benign_correction) { 1795df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1796a3df083aSStefano Zampini PetscScalar *marr; 1797a3df083aSStefano Zampini 1798a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1799a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1800a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1801df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1802a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1803a3df083aSStefano Zampini } 1804a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1805a3df083aSStefano Zampini } 180606656605SStefano Zampini } else { 180706656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 180806656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1809ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1810ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 181106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 181206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 181306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 181406656605SStefano Zampini } 181506656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 181606656605SStefano Zampini } 181780677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1818ffd830a3SStefano Zampini /* S_VV and S_CV */ 181906656605SStefano Zampini if (n_constraints) { 182006656605SStefano Zampini Mat B; 182180677318SStefano Zampini 1822ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 182380677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1824ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1825ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 182680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 182880677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 182980677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 183080677318SStefano Zampini } 1831ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 183280677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 183380677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1834ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 183580677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 183606656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1837ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1838ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 183906656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 184006656605SStefano Zampini } 184104708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 184204708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 184304708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 184404708bb6SStefano Zampini } 1845ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1846ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1847ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1848ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1849ffd830a3SStefano Zampini } 185006656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 185114393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 185214393ed6SStefano Zampini if (need_benign_correction) { 1853df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 185414393ed6SStefano Zampini PetscScalar *marr,*sums; 185514393ed6SStefano Zampini 185614393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 185714393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1858df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 185914393ed6SStefano Zampini const PetscScalar *vals; 186014393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 186114393ed6SStefano Zampini PetscInt n,j,nz; 186214393ed6SStefano Zampini 1863df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1864df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186514393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 186614393ed6SStefano Zampini PetscInt k; 186714393ed6SStefano Zampini sums[j] = 0.; 186814393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 186914393ed6SStefano Zampini } 187014393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 187114393ed6SStefano Zampini for (j=0;j<n;j++) { 187214393ed6SStefano Zampini PetscScalar val = vals[j]; 187314393ed6SStefano Zampini PetscInt k; 187414393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 187514393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 187614393ed6SStefano Zampini } 187714393ed6SStefano Zampini } 187814393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1879df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 188014393ed6SStefano Zampini } 188114393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 188214393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 188314393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 188414393ed6SStefano Zampini } 188580677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 188606656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 188706656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 188806656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 188906656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 189006656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 189106656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 189206656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1893d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1894019a44ceSStefano Zampini } else { 1895d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1896d16cbb6bSStefano Zampini } 18974f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1898019a44ceSStefano Zampini const PetscScalar *vals; 1899019a44ceSStefano Zampini const PetscInt *idxs; 19004f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1901019a44ceSStefano Zampini 19024f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 19034f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1904d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 19054f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 19064f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1907019a44ceSStefano Zampini } 19084f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 190916f15bc4SStefano Zampini } 191021eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1911d16cbb6bSStefano Zampini 191206656605SStefano Zampini /* coarse basis functions */ 191306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 191416f15bc4SStefano Zampini PetscScalar *y; 191516f15bc4SStefano Zampini 1916ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 191706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 191806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 191906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192106656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 192206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 192306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 192406656605SStefano Zampini 192506656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19264f1b2e48SStefano Zampini PetscInt j; 19274f1b2e48SStefano Zampini 192806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 192906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 193006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 193206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19334f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 193406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 193506656605SStefano Zampini } 193606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 193706656605SStefano Zampini } 193804708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 193904708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 194006656605SStefano Zampini } 194106656605SStefano Zampini 194206656605SStefano Zampini if (n_constraints) { 194306656605SStefano Zampini Mat B; 194406656605SStefano Zampini 1945ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 194606656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 194780677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 194806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 194906656605SStefano Zampini if (n_vertices) { 195080677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 195180677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 195280677318SStefano Zampini } else { 195380677318SStefano Zampini Mat S_VCt; 195480677318SStefano Zampini 1955ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1956ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1957ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1958ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1959ffd830a3SStefano Zampini } 196080677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 196180677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 196280677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 196380677318SStefano Zampini } 196406656605SStefano Zampini } 196506656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 196606656605SStefano Zampini /* coarse basis functions */ 196706656605SStefano Zampini for (i=0;i<n_constraints;i++) { 196806656605SStefano Zampini PetscScalar *y; 196906656605SStefano Zampini 1970ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 197106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 197206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 197306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 197606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 197706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19784f1b2e48SStefano Zampini PetscInt j; 19794f1b2e48SStefano Zampini 198006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 198106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 198206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19854f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 198606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 198706656605SStefano Zampini } 198806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 198906656605SStefano Zampini } 199006656605SStefano Zampini } 199180677318SStefano Zampini if (n_constraints) { 199280677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 199380677318SStefano Zampini } 19944f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 1995019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1996019a44ceSStefano Zampini 199706656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 19983301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 1999ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2000ffd830a3SStefano Zampini PetscScalar *marray; 200106656605SStefano Zampini 200206656605SStefano Zampini if (n_constraints) { 2003ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 200406656605SStefano Zampini 2005ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 200606656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2007ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 200816f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 200906656605SStefano Zampini if (n_vertices) { 2010ffd830a3SStefano Zampini Mat S_VCT; 201106656605SStefano Zampini 201206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2013ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 201416f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 201506656605SStefano Zampini } 2016ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 201706656605SStefano Zampini } 201816f15bc4SStefano Zampini if (n_vertices && n_R) { 2019ffd830a3SStefano Zampini PetscScalar *av,*marray; 2020ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2021ffd830a3SStefano Zampini PetscInt n; 2022ffd830a3SStefano Zampini PetscBool flg_row; 202306656605SStefano Zampini 2024ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2025ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2026ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2027ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2028ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2029ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2030ffd830a3SStefano Zampini PetscInt j; 2031ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2032ffd830a3SStefano Zampini } 2033ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2034ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2035ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 203606656605SStefano Zampini } 203706656605SStefano Zampini 2038ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2039ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2040ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2041ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2042ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 204306656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 204406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 204506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 204606656605SStefano Zampini } 2047ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2048ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2049ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2050ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2051ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2052ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2053ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2054ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 205506656605SStefano Zampini } 2056ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 205706656605SStefano Zampini /* coarse basis functions */ 205806656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 205906656605SStefano Zampini PetscScalar *y; 206006656605SStefano Zampini 2061ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 206206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 206306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 206406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206606656605SStefano Zampini if (i<n_vertices) { 206706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 206806656605SStefano Zampini } 206906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 207006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 207106656605SStefano Zampini 207206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 207306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 207406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 207506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 207806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 207906656605SStefano Zampini } 208006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 208106656605SStefano Zampini } 2082ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2083ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 208406656605SStefano Zampini } 2085d62866d3SStefano Zampini /* free memory */ 208688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 208706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 208806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 208906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 209006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2091d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2092d62866d3SStefano Zampini if (n_vertices) { 2093d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2094d62866d3SStefano Zampini } 2095d62866d3SStefano Zampini if (n_constraints) { 2096d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2097d62866d3SStefano Zampini } 209888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 209988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 210088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2101d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 210288ebb749SStefano Zampini Mat coarse_sub_mat; 210325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 210488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 210588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 210688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 21078bec7fa6SStefano Zampini Mat C_B,CPHI; 21088bec7fa6SStefano Zampini IS is_dummy; 21098bec7fa6SStefano Zampini Vec mones; 211088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 211188ebb749SStefano Zampini PetscReal real_value; 211288ebb749SStefano Zampini 2113a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2114a3df083aSStefano Zampini Mat A; 2115a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2116a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2117a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2118a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2119a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2120a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2121a3df083aSStefano Zampini } else { 212288ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 212388ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 212488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 212588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2126a3df083aSStefano Zampini } 212788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 212888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2129ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 213088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 213188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 213288ebb749SStefano Zampini } 213388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 213488ebb749SStefano Zampini 213525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 21363301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 213725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2138ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 213988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 214088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 214188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 214288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 214388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 214488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 214588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 214688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 214788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 214888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 214988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 215088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 215188ebb749SStefano Zampini } else { 215288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 215388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 215488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 215588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 215688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 215788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 215888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 215988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216088ebb749SStefano Zampini } 216188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 216288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 216388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 216488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 21654f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2166d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 2167d12edf2fSStefano Zampini PetscScalar *data,*data2; 21684f1b2e48SStefano Zampini PetscInt j; 2169d12edf2fSStefano Zampini 21704f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 21714f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 21724f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 2173d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2174d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2175d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2176d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 21774f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 21784f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2179d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 21804f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 21814f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 21824f1b2e48SStefano Zampini } 2183d12edf2fSStefano Zampini } 2184d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2185d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2186d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2187d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2188d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2189d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 2190d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 2191d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2192d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 2193d12edf2fSStefano Zampini } 2194d12edf2fSStefano Zampini #if 0 2195d12edf2fSStefano Zampini { 2196d12edf2fSStefano Zampini PetscViewer viewer; 2197d12edf2fSStefano Zampini char filename[256]; 2198ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2199d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2200d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2201ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2202ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2203ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2204d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 2205ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2206ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 2207ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2208ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2209ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2210ffd830a3SStefano Zampini } 2211ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2212ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2213ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2214ffd830a3SStefano Zampini } 2215ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 2216ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2217ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2218ffd830a3SStefano Zampini } 2219d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2220d12edf2fSStefano Zampini } 2221d12edf2fSStefano Zampini #endif 222281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 22238bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 22241575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 222506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 22268bec7fa6SStefano Zampini 22278bec7fa6SStefano Zampini /* check constraints */ 22284f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 22298bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 22308bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 22318bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 22328bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 22338bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 22348bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 22358bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2236bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2237ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2238bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2239bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2240bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2241bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2242bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 224388ebb749SStefano Zampini } 22448bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 22458bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 22468bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 22478bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2248d12edf2fSStefano Zampini } 224925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 225088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 225188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 225288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 225388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 225488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 225588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 225688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 225788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 225888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 225988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2260ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 226288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 226388ebb749SStefano Zampini } 226488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 226588ebb749SStefano Zampini } 22668629588bSStefano Zampini /* get back data */ 22678629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 226888ebb749SStefano Zampini PetscFunctionReturn(0); 226988ebb749SStefano Zampini } 227088ebb749SStefano Zampini 227188ebb749SStefano Zampini #undef __FUNCT__ 2272d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2273d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2274aa0d41d4SStefano Zampini { 2275d65f70fdSStefano Zampini Mat *work_mat; 2276d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2277d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2278d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2279aa0d41d4SStefano Zampini PetscErrorCode ierr; 2280aa0d41d4SStefano Zampini 2281aa0d41d4SStefano Zampini PetscFunctionBegin; 2282d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2283d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2284d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2285d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2286aa0d41d4SStefano Zampini 2287d65f70fdSStefano Zampini if (!rsorted) { 2288906d46d4SStefano Zampini const PetscInt *idxs; 2289906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2290aa0d41d4SStefano Zampini 2291d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2292d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2293d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2294d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2295aa0d41d4SStefano Zampini } 2296d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2297d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2298d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2299d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2300aa0d41d4SStefano Zampini } 2301d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2302d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2303d65f70fdSStefano Zampini } else { 2304d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2305d65f70fdSStefano Zampini isrow_s = isrow; 2306aa0d41d4SStefano Zampini } 2307906d46d4SStefano Zampini 2308d65f70fdSStefano Zampini if (!csorted) { 2309d65f70fdSStefano Zampini if (isrow == iscol) { 2310d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2311d65f70fdSStefano Zampini iscol_s = isrow_s; 2312d65f70fdSStefano Zampini } else { 2313d65f70fdSStefano Zampini const PetscInt *idxs; 2314d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2315906d46d4SStefano Zampini 2316d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2317d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2318d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2319d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2320d65f70fdSStefano Zampini } 2321d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2322d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2323d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2324d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2325d65f70fdSStefano Zampini } 2326d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2327d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2328d65f70fdSStefano Zampini } 2329d65f70fdSStefano Zampini } else { 2330d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2331d65f70fdSStefano Zampini iscol_s = iscol; 2332d65f70fdSStefano Zampini } 2333d65f70fdSStefano Zampini 2334d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2335d65f70fdSStefano Zampini 2336d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2337906d46d4SStefano Zampini Mat new_mat; 2338d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2339906d46d4SStefano Zampini 2340d65f70fdSStefano Zampini if (!rsorted) { 2341d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2342d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2343d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2344d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2345906d46d4SStefano Zampini } 2346d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2347d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2348d65f70fdSStefano Zampini } else { 2349d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2350906d46d4SStefano Zampini } 2351d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2352d65f70fdSStefano Zampini 2353d65f70fdSStefano Zampini if (!csorted) { 2354d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2355d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2356d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2357d65f70fdSStefano Zampini } else { 2358d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2359d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2360d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2361d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2362d65f70fdSStefano Zampini } 2363d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2364d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2365d65f70fdSStefano Zampini } 2366d65f70fdSStefano Zampini } else { 2367d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2368d65f70fdSStefano Zampini } 2369d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2370d65f70fdSStefano Zampini 2371d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2372d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2373d65f70fdSStefano Zampini work_mat[0] = new_mat; 2374d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2375d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2376d65f70fdSStefano Zampini } 2377d65f70fdSStefano Zampini 2378d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2379d65f70fdSStefano Zampini *B = work_mat[0]; 2380d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2381d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2382d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2383d65f70fdSStefano Zampini PetscFunctionReturn(0); 2384d65f70fdSStefano Zampini } 2385d65f70fdSStefano Zampini 2386d65f70fdSStefano Zampini #undef __FUNCT__ 23875e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 23885e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2389aa0d41d4SStefano Zampini { 2390aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 23915e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2392d65f70fdSStefano Zampini Mat new_mat; 23935e8657edSStefano Zampini IS is_local,is_global; 2394d65f70fdSStefano Zampini PetscInt local_size; 2395d65f70fdSStefano Zampini PetscBool isseqaij; 2396aa0d41d4SStefano Zampini PetscErrorCode ierr; 2397aa0d41d4SStefano Zampini 2398aa0d41d4SStefano Zampini PetscFunctionBegin; 2399aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 24005e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 24015e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2402b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2403aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2404d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2405aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2406906d46d4SStefano Zampini 2407906d46d4SStefano Zampini /* check */ 2408906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2409906d46d4SStefano Zampini Vec x,x_change; 2410906d46d4SStefano Zampini PetscReal error; 2411906d46d4SStefano Zampini 24125e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2413906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 24145e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2415e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2416e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2417d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2418e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2419e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2420906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2421906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2422906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2423906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2424906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2425906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2426906d46d4SStefano Zampini } 2427906d46d4SStefano Zampini 242822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 24299b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 243022d5777bSStefano Zampini if (isseqaij) { 24311cf9b237SStefano Zampini Mat M; 24321cf9b237SStefano Zampini 24331cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24341cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24351cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2436aa0d41d4SStefano Zampini } else { 24371cf9b237SStefano Zampini Mat work_mat,M; 24381cf9b237SStefano Zampini 2439aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 24401cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24411cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24421cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2443aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2444aa0d41d4SStefano Zampini } 24453301b35fSStefano Zampini if (matis->A->symmetric_set) { 24463301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2447e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24483301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2449e496cd5dSStefano Zampini #endif 24503301b35fSStefano Zampini } 245145a1bb75SStefano Zampini /* 245245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2453d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 245445a1bb75SStefano Zampini */ 2455d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2456aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2457aa0d41d4SStefano Zampini } 2458aa0d41d4SStefano Zampini 2459aa0d41d4SStefano Zampini #undef __FUNCT__ 2460a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 24618ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2462a64d13efSStefano Zampini { 2463a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2464a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2465d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 246653892102SStefano Zampini PetscInt *idx_R_local=NULL; 24673a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 24683a50541eSStefano Zampini PetscInt vbs,bs; 24696816873aSStefano Zampini PetscBT bitmask=NULL; 2470a64d13efSStefano Zampini PetscErrorCode ierr; 2471a64d13efSStefano Zampini 2472a64d13efSStefano Zampini PetscFunctionBegin; 2473b23d619eSStefano Zampini /* 2474b23d619eSStefano Zampini No need to setup local scatters if 2475b23d619eSStefano Zampini - primal space is unchanged 2476b23d619eSStefano Zampini AND 2477b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2478b23d619eSStefano Zampini AND 2479b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2480b23d619eSStefano Zampini */ 2481b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2482f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2483f4ddd8eeSStefano Zampini } 2484f4ddd8eeSStefano Zampini /* destroy old objects */ 2485f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2486f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2487f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2488a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2489b371cd4fSStefano Zampini n_B = pcis->n_B; 2490b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2491b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 24923a50541eSStefano Zampini 2493a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 24946816873aSStefano Zampini 249553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2496df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2497854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2498a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2499a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25000e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2501a64d13efSStefano Zampini } 2502a64d13efSStefano Zampini 2503a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 25044641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 25056816873aSStefano Zampini idx_R_local[n_R++] = i; 2506a64d13efSStefano Zampini } 2507a64d13efSStefano Zampini } 2508df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2509df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 25106816873aSStefano Zampini 2511df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2512df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 25136816873aSStefano Zampini } 25143a50541eSStefano Zampini 25153a50541eSStefano Zampini /* Block code */ 25163a50541eSStefano Zampini vbs = 1; 25173a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 25183a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 25193a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 25203a50541eSStefano Zampini PetscInt *vary; 2521df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2522785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 25233a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2524d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2525d3df7717SStefano 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 */ 25260e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2527d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 25283a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 25293a50541eSStefano Zampini is_blocked = PETSC_FALSE; 25303a50541eSStefano Zampini break; 25313a50541eSStefano Zampini } 25323a50541eSStefano Zampini } 2533d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2534d3df7717SStefano Zampini } else { 2535d3df7717SStefano Zampini /* Verify directly the R set */ 2536d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2537d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2538d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2539d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2540d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2541d3df7717SStefano Zampini break; 2542d3df7717SStefano Zampini } 2543d3df7717SStefano Zampini } 2544d3df7717SStefano Zampini } 2545d3df7717SStefano Zampini } 25463a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 25473a50541eSStefano Zampini vbs = bs; 25483a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 25493a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 25503a50541eSStefano Zampini } 25513a50541eSStefano Zampini } 25523a50541eSStefano Zampini } 25533a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2554df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2555df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 255653892102SStefano Zampini 2557df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2558df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 255953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2560df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 256153892102SStefano Zampini } else { 25623a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 256353892102SStefano Zampini } 2564a64d13efSStefano Zampini 2565a64d13efSStefano Zampini /* print some info if requested */ 2566a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2567a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2568a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 25691575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2570a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2571a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 25724f1b2e48SStefano 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); 2573a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2574a64d13efSStefano Zampini } 2575a64d13efSStefano Zampini 2576a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2577df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 25786816873aSStefano Zampini IS is_aux1,is_aux2; 25796816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 25806816873aSStefano Zampini 25813a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2582854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2583854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2584a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 25854641a718SStefano Zampini for (i=0; i<n_D; i++) { 25864641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 25874641a718SStefano Zampini } 2588a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2589a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 25904641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 25914641a718SStefano Zampini aux_array1[j++] = i; 2592a64d13efSStefano Zampini } 2593a64d13efSStefano Zampini } 2594a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2595a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2596a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 25974641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 25984641a718SStefano Zampini aux_array2[j++] = i; 2599a64d13efSStefano Zampini } 2600a64d13efSStefano Zampini } 2601a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2602a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2603a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2604a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2605a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2606a64d13efSStefano Zampini 26078eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2608785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2609a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26104641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 26114641a718SStefano Zampini aux_array1[j++] = i; 2612a64d13efSStefano Zampini } 2613a64d13efSStefano Zampini } 2614a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2615a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2616a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2617a64d13efSStefano Zampini } 26184641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 26193a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2620d62866d3SStefano Zampini } else { 2621df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26226816873aSStefano Zampini IS tis; 26236816873aSStefano Zampini PetscInt schur_size; 26246816873aSStefano Zampini 2625df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 26266816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2627df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 26286816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 26296816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 26306816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 26316816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 26326816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2633d62866d3SStefano Zampini } 2634d62866d3SStefano Zampini } 2635a64d13efSStefano Zampini PetscFunctionReturn(0); 2636a64d13efSStefano Zampini } 2637a64d13efSStefano Zampini 2638304d26faSStefano Zampini 2639304d26faSStefano Zampini #undef __FUNCT__ 2640304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2641684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2642304d26faSStefano Zampini { 2643304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2644304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2645304d26faSStefano Zampini PC pc_temp; 2646304d26faSStefano Zampini Mat A_RR; 2647f4ddd8eeSStefano Zampini MatReuse reuse; 2648304d26faSStefano Zampini PetscScalar m_one = -1.0; 2649304d26faSStefano Zampini PetscReal value; 265004708bb6SStefano Zampini PetscInt n_D,n_R; 26519577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2652304d26faSStefano Zampini PetscErrorCode ierr; 2653e604994aSStefano Zampini /* prefixes stuff */ 2654312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2655e604994aSStefano Zampini size_t len; 2656304d26faSStefano Zampini 2657304d26faSStefano Zampini PetscFunctionBegin; 2658304d26faSStefano Zampini 2659e604994aSStefano Zampini /* compute prefixes */ 2660e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2661e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2662e604994aSStefano Zampini if (!pcbddc->current_level) { 2663e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2664e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2665e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2666e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2667e604994aSStefano Zampini } else { 2668e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2669312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2670e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2671e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2672312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2673312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 267434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 267534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2676e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2677e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2678e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2679e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2680e604994aSStefano Zampini } 2681e604994aSStefano Zampini 2682304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2683684f6988SStefano Zampini if (dirichlet) { 2684d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2685a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2686a3df083aSStefano Zampini Mat A_IIn; 2687a3df083aSStefano Zampini 2688a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2689a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2690a3df083aSStefano Zampini pcis->A_II = A_IIn; 2691a3df083aSStefano Zampini } 26923301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 26933301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2694964fefecSStefano Zampini } 2695ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2696964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2697304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2698304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2699304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2700304d26faSStefano Zampini /* default */ 2701304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2702e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 27039577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2704304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 27059577ea80SStefano Zampini if (issbaij) { 27069577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 27079577ea80SStefano Zampini } else { 2708304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 27099577ea80SStefano Zampini } 2710304d26faSStefano Zampini /* Allow user's customization */ 2711304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2712304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2713304d26faSStefano Zampini } 2714d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2715df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2716df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2717d62866d3SStefano Zampini 2718df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2719d5574798SStefano Zampini } 2720304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2721304d26faSStefano Zampini if (!n_D) { 2722304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2723304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2724304d26faSStefano Zampini } 2725304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2726304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2727304d26faSStefano Zampini /* set ksp_D into pcis data */ 2728304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2729304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2730304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2731684f6988SStefano Zampini } 2732304d26faSStefano Zampini 2733304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2734684f6988SStefano Zampini A_RR = 0; 2735684f6988SStefano Zampini if (neumann) { 2736d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 273704708bb6SStefano Zampini PetscInt ibs,mbs; 273804708bb6SStefano Zampini PetscBool issbaij; 273904708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2740f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 27418ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2742f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2743f4ddd8eeSStefano Zampini PetscInt nn_R; 274481d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2745f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2746f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2747f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2748f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2749f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2750f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2751f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2752727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2753f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2754f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2755f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2756f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2757f4ddd8eeSStefano Zampini } 2758f4ddd8eeSStefano Zampini } 2759f4ddd8eeSStefano Zampini /* last check */ 2760d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2761f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2762f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2763f4ddd8eeSStefano Zampini } 2764f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2765f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2766f4ddd8eeSStefano Zampini } 2767f4ddd8eeSStefano Zampini /* extract A_RR */ 2768af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2769af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 277004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 277104708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 277204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 277304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 277404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2775af732b37SStefano Zampini } else { 277604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 27776816873aSStefano Zampini } 277804708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 277904708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 278004708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 278104708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 278204708bb6SStefano Zampini } else { 278304708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 278404708bb6SStefano Zampini } 278504708bb6SStefano Zampini } 2786f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 27873301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27883301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 27896816873aSStefano Zampini } 2790a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2791a3df083aSStefano Zampini Mat A_RRn; 2792a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RRn);CHKERRQ(ierr); 2793a3df083aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2794a3df083aSStefano Zampini A_RR = A_RRn; 2795a3df083aSStefano Zampini } 2796f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2797304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2798304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2799304d26faSStefano Zampini /* default */ 2800304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2801e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2802304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 28039577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 28049577ea80SStefano Zampini if (issbaij) { 28059577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28069577ea80SStefano Zampini } else { 2807304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28089577ea80SStefano Zampini } 2809304d26faSStefano Zampini /* Allow user's customization */ 2810304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2811304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2812304d26faSStefano Zampini } 2813304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2814304d26faSStefano Zampini if (!n_R) { 2815304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2816304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2817304d26faSStefano Zampini } 2818df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2819df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2820df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2821d62866d3SStefano Zampini 2822df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2823d62866d3SStefano Zampini } 2824ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2825304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2826304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2827684f6988SStefano Zampini } 28286816873aSStefano Zampini /* free Neumann problem's matrix */ 28296816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2830304d26faSStefano Zampini 2831304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 28320fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2833684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2834684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28351575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2836684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2837684f6988SStefano Zampini } 2838684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 28390fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 28400fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 28410fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 28420fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 28430fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2844304d26faSStefano Zampini /* need to be adapted? */ 2845b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2846b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2847b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2848304d26faSStefano Zampini /* print info */ 2849304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2850e604994aSStefano 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); 2851304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2852304d26faSStefano Zampini } 2853b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2854298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2855304d26faSStefano Zampini } 2856684f6988SStefano Zampini } 2857684f6988SStefano Zampini if (neumann) { /* Neumann */ 28586816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 28590fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 28600fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 28610fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 28620fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 28630fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2864304d26faSStefano Zampini /* need to be adapted? */ 2865b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2866b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2867304d26faSStefano Zampini /* print info */ 2868304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2869e604994aSStefano 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); 2870304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2871304d26faSStefano Zampini } 2872b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2873298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2874304d26faSStefano Zampini } 28750fccc4e9SStefano Zampini } 2876684f6988SStefano Zampini } 2877304d26faSStefano Zampini PetscFunctionReturn(0); 2878304d26faSStefano Zampini } 2879304d26faSStefano Zampini 2880304d26faSStefano Zampini #undef __FUNCT__ 2881ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 288280677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2883674ae819SStefano Zampini { 2884674ae819SStefano Zampini PetscErrorCode ierr; 2885674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2886be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2887674ae819SStefano Zampini 2888674ae819SStefano Zampini PetscFunctionBegin; 2889df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 289080677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 289120c7b377SStefano Zampini } 289280677318SStefano Zampini if (!pcbddc->switch_static) { 289380677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 289480677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 289580677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 289620c7b377SStefano Zampini } 2897df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 289880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 289980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 290020c7b377SStefano Zampini } else { 2901df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2902be83ff47SStefano Zampini 2903df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2904df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 290520c7b377SStefano Zampini } 2906be83ff47SStefano Zampini } else { 290780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 290880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 290980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 291080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 291180677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 291280677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 291380677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 291480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 291580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2916674ae819SStefano Zampini } 2917674ae819SStefano Zampini } 2918df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 291980677318SStefano Zampini if (applytranspose) { 292080677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 292180677318SStefano Zampini } else { 292280677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 292380677318SStefano Zampini } 2924be83ff47SStefano Zampini } else { 2925df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2926be83ff47SStefano Zampini 2927be83ff47SStefano Zampini if (applytranspose) { 2928df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2929be83ff47SStefano Zampini } else { 2930df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2931be83ff47SStefano Zampini } 2932be83ff47SStefano Zampini } 293380677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 293480677318SStefano Zampini if (!pcbddc->switch_static) { 2935df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 293680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 293780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2938be83ff47SStefano Zampini } else { 2939df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2940be83ff47SStefano Zampini 2941df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2942df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2943be83ff47SStefano Zampini } 294480677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 294580677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 294680677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 294780677318SStefano Zampini } 294880677318SStefano Zampini } else { 294980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 295480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 295580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 295680677318SStefano Zampini } 295780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 296080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2961674ae819SStefano Zampini } 2962674ae819SStefano Zampini PetscFunctionReturn(0); 2963674ae819SStefano Zampini } 2964674ae819SStefano Zampini 2965dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2966674ae819SStefano Zampini #undef __FUNCT__ 2967674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2968dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2969674ae819SStefano Zampini { 2970674ae819SStefano Zampini PetscErrorCode ierr; 2971674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2972674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2973674ae819SStefano Zampini const PetscScalar zero = 0.0; 2974674ae819SStefano Zampini 2975674ae819SStefano Zampini PetscFunctionBegin; 2976dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2977dc359a40SStefano Zampini if (applytranspose) { 2978674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 29798eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2980dc359a40SStefano Zampini } else { 2981674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2982674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 298315aaf578SStefano Zampini } 2984efc2fbd9SStefano Zampini 2985efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 29864f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2987efc2fbd9SStefano Zampini PetscScalar *array; 29884f1b2e48SStefano Zampini PetscInt j; 2989efc2fbd9SStefano Zampini 2990efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 29914f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 2992efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2993efc2fbd9SStefano Zampini } 2994efc2fbd9SStefano Zampini 299512edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 299612edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 299712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299912edc857SStefano Zampini 30009f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 300112edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 300212edc857SStefano Zampini if (pcbddc->coarse_ksp) { 300351694757SStefano Zampini Mat coarse_mat; 3004964fefecSStefano Zampini Vec rhs,sol; 300551694757SStefano Zampini MatNullSpace nullsp; 3006964fefecSStefano Zampini 3007964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3008964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 300951694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 301051694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 301151694757SStefano Zampini if (nullsp) { 301251694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 301351694757SStefano Zampini } 301412edc857SStefano Zampini if (applytranspose) { 3015964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 301612edc857SStefano Zampini } else { 3017964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 301812edc857SStefano Zampini } 301951694757SStefano Zampini if (nullsp) { 302051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 302151694757SStefano Zampini } 302212edc857SStefano Zampini } 3023674ae819SStefano Zampini 3024674ae819SStefano Zampini /* Local solution on R nodes */ 302580677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 302680677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 30279f00e9b4SStefano Zampini } 3028674ae819SStefano Zampini 30299f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 30309f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 303112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3032674ae819SStefano Zampini 3033674ae819SStefano Zampini /* Sum contributions from two levels */ 3034dc359a40SStefano Zampini if (applytranspose) { 3035dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3036dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3037dc359a40SStefano Zampini } else { 3038674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 30398eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3040dc359a40SStefano Zampini } 3041efc2fbd9SStefano Zampini /* store p0 */ 30424f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3043efc2fbd9SStefano Zampini PetscScalar *array; 30444f1b2e48SStefano Zampini PetscInt j; 3045efc2fbd9SStefano Zampini 3046efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30474f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3048efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3049efc2fbd9SStefano Zampini } 3050674ae819SStefano Zampini PetscFunctionReturn(0); 3051674ae819SStefano Zampini } 3052674ae819SStefano Zampini 3053674ae819SStefano Zampini #undef __FUNCT__ 3054674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 305512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3056674ae819SStefano Zampini { 3057674ae819SStefano Zampini PetscErrorCode ierr; 3058674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 305958da7f69SStefano Zampini PetscScalar *array; 306012edc857SStefano Zampini Vec from,to; 3061674ae819SStefano Zampini 3062674ae819SStefano Zampini PetscFunctionBegin; 306312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 306412edc857SStefano Zampini from = pcbddc->coarse_vec; 306512edc857SStefano Zampini to = pcbddc->vec1_P; 306612edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 306712edc857SStefano Zampini Vec tvec; 306858da7f69SStefano Zampini 306958da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 307058da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 307112edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 307258da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 307358da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 307458da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 307512edc857SStefano Zampini } 307612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 307712edc857SStefano Zampini from = pcbddc->vec1_P; 307812edc857SStefano Zampini to = pcbddc->coarse_vec; 307912edc857SStefano Zampini } 308012edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3081674ae819SStefano Zampini PetscFunctionReturn(0); 3082674ae819SStefano Zampini } 3083674ae819SStefano Zampini 3084674ae819SStefano Zampini #undef __FUNCT__ 3085674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 308612edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3087674ae819SStefano Zampini { 3088674ae819SStefano Zampini PetscErrorCode ierr; 3089674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 309058da7f69SStefano Zampini PetscScalar *array; 309112edc857SStefano Zampini Vec from,to; 3092674ae819SStefano Zampini 3093674ae819SStefano Zampini PetscFunctionBegin; 309412edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 309512edc857SStefano Zampini from = pcbddc->coarse_vec; 309612edc857SStefano Zampini to = pcbddc->vec1_P; 309712edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 309812edc857SStefano Zampini from = pcbddc->vec1_P; 309912edc857SStefano Zampini to = pcbddc->coarse_vec; 310012edc857SStefano Zampini } 310112edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 310212edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 310312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 310412edc857SStefano Zampini Vec tvec; 310558da7f69SStefano Zampini 310612edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 310758da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 310858da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 310958da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 311058da7f69SStefano Zampini } 311158da7f69SStefano Zampini } else { 311258da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 311358da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 311412edc857SStefano Zampini } 311512edc857SStefano Zampini } 3116674ae819SStefano Zampini PetscFunctionReturn(0); 3117674ae819SStefano Zampini } 3118674ae819SStefano Zampini 3119984c4197SStefano Zampini /* uncomment for testing purposes */ 3120984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3121674ae819SStefano Zampini #undef __FUNCT__ 3122674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3123674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3124674ae819SStefano Zampini { 3125674ae819SStefano Zampini PetscErrorCode ierr; 3126674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3127674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3128674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3129984c4197SStefano Zampini /* one and zero */ 3130984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3131984c4197SStefano Zampini /* space to store constraints and their local indices */ 31329162d606SStefano Zampini PetscScalar *constraints_data; 31339162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 31349162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 31359162d606SStefano Zampini PetscInt *constraints_n; 3136984c4197SStefano Zampini /* iterators */ 3137b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3138984c4197SStefano Zampini /* BLAS integers */ 3139e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3140e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3141c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3142727cdba6SStefano Zampini /* reuse */ 31430e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 31440e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3145984c4197SStefano Zampini /* change of basis */ 3146b3d85658SStefano Zampini PetscBool qr_needed; 31479162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3148984c4197SStefano Zampini /* auxiliary stuff */ 314964efe560SStefano Zampini PetscInt *nnz,*is_indices; 31508a0068c3SStefano Zampini PetscInt ncc; 3151984c4197SStefano Zampini /* some quantities */ 315245a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3153a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3154984c4197SStefano Zampini 3155674ae819SStefano Zampini PetscFunctionBegin; 31568e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 31578e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 31588e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3159088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3160088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 31610e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 31620e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 31630e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 31640e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 31650e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3166088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3167cf5a6209SStefano Zampini 3168cf5a6209SStefano Zampini /* print some info */ 3169cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3170cf5a6209SStefano Zampini IS vertices; 3171cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3172cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3173cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3174cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 31751575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3176cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3177cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3178fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3179fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3180cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31811575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3182cf5a6209SStefano Zampini } 3183cf5a6209SStefano Zampini 3184cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 31859162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3186cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3187cf5a6209SStefano Zampini const Vec *nearnullvecs; 3188cf5a6209SStefano Zampini Vec *localnearnullsp; 3189cf5a6209SStefano Zampini PetscScalar *array; 3190cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3191cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3192674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3193b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3194674ae819SStefano Zampini PetscScalar *work; 3195674ae819SStefano Zampini PetscReal *singular_vals; 3196674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3197674ae819SStefano Zampini PetscReal *rwork; 3198674ae819SStefano Zampini #endif 3199674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3200674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3201674ae819SStefano Zampini #else 3202964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3203964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3204674ae819SStefano Zampini #endif 3205674ae819SStefano Zampini 3206674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3207d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3208d06fc5fdSStefano Zampini /* free unneeded index sets */ 3209d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3210d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3211674ae819SStefano Zampini } 3212d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3213d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3214d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3215d06fc5fdSStefano Zampini } 3216d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3217d06fc5fdSStefano Zampini n_ISForEdges = 0; 3218d06fc5fdSStefano Zampini } 3219d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3220d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3221d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3222d06fc5fdSStefano Zampini } 3223d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3224d06fc5fdSStefano Zampini n_ISForFaces = 0; 3225d06fc5fdSStefano Zampini } 322670022509SStefano Zampini 322770022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 322870022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 322970022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 323070022509SStefano Zampini if (pcbddc->NullSpace) { 323170022509SStefano Zampini PetscBool tbool[2],gbool[2]; 323270022509SStefano Zampini 323370022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3234b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3235d06fc5fdSStefano Zampini if (!ISForEdges) { 3236d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3237d06fc5fdSStefano Zampini } 3238b8ffe317SStefano Zampini } 3239d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3240d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3241d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3242d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3243d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 324498a51de6SStefano Zampini } 324570022509SStefano Zampini #endif 324608122e43SStefano Zampini 3247674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3248674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3249674ae819SStefano Zampini if (nearnullsp) { 3250674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3251f4ddd8eeSStefano Zampini /* remove any stored info */ 3252f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3253f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3254f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3255f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3256f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3257473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3258f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3259f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3260f4ddd8eeSStefano Zampini } 3261984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3262984c4197SStefano Zampini nnsp_size = 0; 3263674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3264674ae819SStefano Zampini } 3265984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3266984c4197SStefano Zampini max_constraints = nnsp_size; 3267984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3268984c4197SStefano Zampini 3269674ae819SStefano Zampini /* 3270674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 32719162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 32729162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 32739162d606SStefano Zampini There can be multiple constraints per connected component 3274674ae819SStefano Zampini */ 3275674ae819SStefano Zampini n_vertices = 0; 3276674ae819SStefano Zampini if (ISForVertices) { 3277674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3278674ae819SStefano Zampini } 32799162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 32809162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 32819162d606SStefano Zampini 32829162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 32839162d606SStefano Zampini total_counts *= max_constraints; 3284674ae819SStefano Zampini total_counts += n_vertices; 32854641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 32869162d606SStefano Zampini 3287674ae819SStefano Zampini total_counts = 0; 3288674ae819SStefano Zampini max_size_of_constraint = 0; 3289674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 32909162d606SStefano Zampini IS used_is; 3291674ae819SStefano Zampini if (i<n_ISForEdges) { 32929162d606SStefano Zampini used_is = ISForEdges[i]; 3293674ae819SStefano Zampini } else { 32949162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3295674ae819SStefano Zampini } 32969162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3297674ae819SStefano Zampini total_counts += j; 3298674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3299674ae819SStefano Zampini } 33009162d606SStefano 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); 33019162d606SStefano Zampini 3302984c4197SStefano Zampini /* get local part of global near null space vectors */ 3303785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3304984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3305984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3306e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3307e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3308984c4197SStefano Zampini } 3309674ae819SStefano Zampini 3310242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3311242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3312a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3313242a89d7SStefano Zampini 3314984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3315a773dcb8SStefano Zampini if (!skip_lapack) { 3316674ae819SStefano Zampini PetscScalar temp_work; 3317911cabfeSStefano Zampini 3318674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3319984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3320785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3321785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3322785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3323674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3324785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3325674ae819SStefano Zampini #endif 3326674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3327c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3328c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3329674ae819SStefano Zampini lwork = -1; 3330674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3331674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3332c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3333674ae819SStefano Zampini #else 3334c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3335674ae819SStefano Zampini #endif 3336674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3337984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3338674ae819SStefano Zampini #else /* on missing GESVD */ 3339674ae819SStefano Zampini /* SVD */ 3340674ae819SStefano Zampini PetscInt max_n,min_n; 3341674ae819SStefano Zampini max_n = max_size_of_constraint; 3342984c4197SStefano Zampini min_n = max_constraints; 3343984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3344674ae819SStefano Zampini min_n = max_size_of_constraint; 3345984c4197SStefano Zampini max_n = max_constraints; 3346674ae819SStefano Zampini } 3347785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3348674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3349785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3350674ae819SStefano Zampini #endif 3351674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3352674ae819SStefano Zampini lwork = -1; 3353e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3354e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3355b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3356674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3357674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 33589162d606SStefano 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)); 3359674ae819SStefano Zampini #else 33609162d606SStefano 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)); 3361674ae819SStefano Zampini #endif 3362674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3363984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3364984c4197SStefano Zampini #endif /* on missing GESVD */ 3365674ae819SStefano Zampini /* Allocate optimal workspace */ 3366674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3367854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3368674ae819SStefano Zampini } 3369674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3370674ae819SStefano Zampini total_counts = 0; 33719162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 33729162d606SStefano Zampini constraints_data_ptr[0] = 0; 3373674ae819SStefano Zampini /* vertices */ 33749162d606SStefano Zampini if (n_vertices) { 3375674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 33769162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3377674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 33789162d606SStefano Zampini constraints_n[total_counts] = 1; 33799162d606SStefano Zampini constraints_data[total_counts] = 1.0; 33809162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 33819162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3382674ae819SStefano Zampini total_counts++; 3383674ae819SStefano Zampini } 3384674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3385674ae819SStefano Zampini n_vertices = total_counts; 3386674ae819SStefano Zampini } 3387984c4197SStefano Zampini 3388674ae819SStefano Zampini /* edges and faces */ 33899162d606SStefano Zampini total_counts_cc = total_counts; 3390911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 33919162d606SStefano Zampini IS used_is; 33929162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 33939162d606SStefano Zampini 3394911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 33959162d606SStefano Zampini used_is = ISForEdges[ncc]; 3396984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3397674ae819SStefano Zampini } else { 33989162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3399984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3400674ae819SStefano Zampini } 3401674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 34029162d606SStefano Zampini 34039162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 34049162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3405984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3406984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3407674ae819SStefano Zampini if (nnsp_has_cnst) { 34085b08dc53SStefano Zampini PetscScalar quad_value; 34099162d606SStefano Zampini 34109162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34119162d606SStefano Zampini idxs_copied = PETSC_TRUE; 34129162d606SStefano Zampini 3413a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3414674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3415a773dcb8SStefano Zampini } else { 3416a773dcb8SStefano Zampini quad_value = 1.0; 3417a773dcb8SStefano Zampini } 3418674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34199162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3420674ae819SStefano Zampini } 34219162d606SStefano Zampini temp_constraints++; 3422674ae819SStefano Zampini total_counts++; 3423674ae819SStefano Zampini } 3424674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3425984c4197SStefano Zampini PetscReal real_value; 34269162d606SStefano Zampini PetscScalar *ptr_to_data; 34279162d606SStefano Zampini 3428984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 34299162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3430674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34319162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3432674ae819SStefano Zampini } 3433984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3434984c4197SStefano Zampini /* check if array is null on the connected component */ 3435e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34369162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 34375b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3438674ae819SStefano Zampini temp_constraints++; 3439674ae819SStefano Zampini total_counts++; 34409162d606SStefano Zampini if (!idxs_copied) { 34419162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34429162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3443674ae819SStefano Zampini } 3444674ae819SStefano Zampini } 34459162d606SStefano Zampini } 34469162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 344745a1bb75SStefano Zampini valid_constraints = temp_constraints; 3448eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3449a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 34509162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 34519162d606SStefano Zampini 34529162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3453a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34549162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3455a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 34569162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3457a773dcb8SStefano Zampini } else { /* perform SVD */ 3458984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 34599162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3460674ae819SStefano Zampini 3461674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3462984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3463984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3464984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3465984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3466984c4197SStefano Zampini from that computed using LAPACKgesvd 3467984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3468984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3469984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3470674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3471e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3472984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3473674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3474674ae819SStefano Zampini for (k=0;k<j+1;k++) { 34759162d606SStefano 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)); 3476674ae819SStefano Zampini } 3477674ae819SStefano Zampini } 3478e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3479e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3480e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3481674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3482c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3483674ae819SStefano Zampini #else 3484c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3485674ae819SStefano Zampini #endif 3486674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3487984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3488984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3489674ae819SStefano Zampini j = 0; 3490984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3491674ae819SStefano Zampini total_counts = total_counts-j; 349245a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3493e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3494c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3495c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3496c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3497c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3498c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3499c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3500674ae819SStefano Zampini if (j<temp_constraints) { 3501984c4197SStefano Zampini PetscInt ii; 3502984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3503674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 35049162d606SStefano 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)); 3505674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3506984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3507674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 35089162d606SStefano 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]; 3509674ae819SStefano Zampini } 3510674ae819SStefano Zampini } 3511674ae819SStefano Zampini } 3512674ae819SStefano Zampini #else /* on missing GESVD */ 3513e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3514e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3515b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3516674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3517674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35189162d606SStefano Zampini PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr)); 3519674ae819SStefano Zampini #else 35209162d606SStefano 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)); 3521674ae819SStefano Zampini #endif 3522984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3523674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3524984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3525e310c8b4SStefano Zampini k = temp_constraints; 3526e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3527674ae819SStefano Zampini j = 0; 3528e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 352945a1bb75SStefano Zampini valid_constraints = k-j; 3530911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3531984c4197SStefano Zampini #endif /* on missing GESVD */ 3532674ae819SStefano Zampini } 3533a773dcb8SStefano Zampini } 35349162d606SStefano Zampini /* update pointers information */ 35359162d606SStefano Zampini if (valid_constraints) { 35369162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 35379162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 35389162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 35399162d606SStefano Zampini /* set change_of_basis flag */ 354045a1bb75SStefano Zampini if (boolforchange) { 3541b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 35429162d606SStefano Zampini } 3543b3d85658SStefano Zampini total_counts_cc++; 354445a1bb75SStefano Zampini } 354545a1bb75SStefano Zampini } 3546984c4197SStefano Zampini /* free workspace */ 35478f1c130eSStefano Zampini if (!skip_lapack) { 3548984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3549984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3550984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3551984c4197SStefano Zampini #endif 3552984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3553984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3554984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3555984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3556984c4197SStefano Zampini #endif 3557984c4197SStefano Zampini } 3558984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3559984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3560984c4197SStefano Zampini } 3561984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3562cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3563cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3564cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3565cf5a6209SStefano Zampini } 3566cf5a6209SStefano Zampini if (n_ISForFaces) { 3567cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3568cf5a6209SStefano Zampini } 3569cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3570cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3571cf5a6209SStefano Zampini } 3572cf5a6209SStefano Zampini if (n_ISForEdges) { 3573cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3574cf5a6209SStefano Zampini } 3575cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 357608122e43SStefano Zampini } else { 357708122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3578984c4197SStefano Zampini 357908122e43SStefano Zampini total_counts = 0; 358008122e43SStefano Zampini n_vertices = 0; 3581d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3582d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 358308122e43SStefano Zampini } 358408122e43SStefano Zampini max_constraints = 0; 35859162d606SStefano Zampini total_counts_cc = 0; 358608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 358708122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 35889162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 358908122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 359008122e43SStefano Zampini } 35919162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 35929162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 35939162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 35949162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 359574d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 35969162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 35979162d606SStefano Zampini total_counts_cc = 0; 35989162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 35999162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 36009162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 360108122e43SStefano Zampini } 360208122e43SStefano Zampini } 36039162d606SStefano Zampini #if 0 36049162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 36059162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36069162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 36079162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 36089162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 36099162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 36109162d606SStefano Zampini } 36119162d606SStefano Zampini printf("\n"); 36129162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 36139162d606SStefano Zampini } 36141b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 36158bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 36161b968477SStefano Zampini } 36171b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 36188bec7fa6SStefano 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]); 36191b968477SStefano Zampini } 362008122e43SStefano Zampini #endif 362108122e43SStefano Zampini 36228bec7fa6SStefano Zampini max_size_of_constraint = 0; 36239162d606SStefano 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]); 36249162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 362508122e43SStefano Zampini /* Change of basis */ 3626b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 362708122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 362808122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 362908122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3630b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 363108122e43SStefano Zampini } 363208122e43SStefano Zampini } 363308122e43SStefano Zampini } 363408122e43SStefano Zampini } 3635984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 36364f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 363708122e43SStefano Zampini 36389162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 36399162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 36409162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 36419162d606SStefano 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); 364208122e43SStefano Zampini } 3643674ae819SStefano Zampini 3644674ae819SStefano Zampini /* Create constraint matrix */ 3645674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 364616f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3647984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3648984c4197SStefano Zampini 3649984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3650a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3651a717540cSStefano Zampini qr_needed = PETSC_FALSE; 365274d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3653984c4197SStefano Zampini total_primal_vertices=0; 3654b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 36559162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36569162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3657984c4197SStefano Zampini if (size_of_constraint == 1) { 36589162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3659b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 366064efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 36619162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 36629162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3663a717540cSStefano Zampini } 3664b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 366574d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3666a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3667a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3668a717540cSStefano Zampini } 3669fa434743SStefano Zampini } else { 3670b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3671fa434743SStefano Zampini } 3672a717540cSStefano Zampini } 3673b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3674b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3675674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 367670022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3677b3d85658SStefano Zampini 36784f1b2e48SStefano 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); 36790e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 36800e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3681984c4197SStefano Zampini 3682984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 368374d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3684785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3685984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 368674d5cdf7SStefano Zampini 3687984c4197SStefano Zampini j = total_primal_vertices; 368874d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3689b3d85658SStefano Zampini cum = total_primal_vertices; 36909162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 36914641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3692b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3693b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3694b3d85658SStefano Zampini cum++; 36959162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 369674d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 369774d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 369874d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 369974d5cdf7SStefano Zampini } 37009162d606SStefano Zampini j += constraints_n[i]; 3701674ae819SStefano Zampini } 3702674ae819SStefano Zampini } 3703674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3704674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3705088faed8SStefano Zampini 3706674ae819SStefano Zampini /* set values in constraint matrix */ 3707984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 37080e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3709674ae819SStefano Zampini } 3710984c4197SStefano Zampini total_counts = total_primal_vertices; 37119162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37124641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 37139162d606SStefano Zampini PetscInt *cols; 37149162d606SStefano Zampini 37159162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 37169162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 37179162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37189162d606SStefano Zampini PetscInt row = total_counts+k; 37199162d606SStefano Zampini PetscScalar *vals; 37209162d606SStefano Zampini 37219162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 37229162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 37239162d606SStefano Zampini } 37249162d606SStefano Zampini total_counts += constraints_n[i]; 3725674ae819SStefano Zampini } 3726674ae819SStefano Zampini } 3727674ae819SStefano Zampini /* assembling */ 3728674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3729674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3730088faed8SStefano Zampini 3731984c4197SStefano Zampini /* 373245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3733984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3734984c4197SStefano Zampini */ 3735674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3736674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3737026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3738984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3739984c4197SStefano Zampini /* working stuff for GEQRF */ 374081d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3741984c4197SStefano Zampini PetscBLASInt lqr_work; 3742984c4197SStefano Zampini /* working stuff for UNGQR */ 3743984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3744984c4197SStefano Zampini PetscBLASInt lgqr_work; 3745984c4197SStefano Zampini /* working stuff for TRTRS */ 3746984c4197SStefano Zampini PetscScalar *trs_rhs; 37473f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3748984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3749984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3750984c4197SStefano Zampini PetscScalar *start_vals; 3751984c4197SStefano Zampini /* working stuff for values insertion */ 37524641a718SStefano Zampini PetscBT is_primal; 375364efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3754906d46d4SStefano Zampini /* matrix sizes */ 3755906d46d4SStefano Zampini PetscInt global_size,local_size; 3756906d46d4SStefano Zampini /* temporary change of basis */ 3757906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3758cf5a6209SStefano Zampini /* extra space for debugging */ 3759cf5a6209SStefano Zampini PetscScalar *dbg_work; 3760984c4197SStefano Zampini 3761906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3762906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 376316f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3764bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3765906d46d4SStefano Zampini /* nonzeros for local mat */ 3766bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3767bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 37689162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3769a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 37709162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3771a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 37729162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3773a717540cSStefano Zampini } else { 37749162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 37759162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3776a717540cSStefano Zampini } 3777a717540cSStefano Zampini } 3778a717540cSStefano Zampini } 3779906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3780bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3781a717540cSStefano Zampini /* Set initial identity in the matrix */ 3782bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3783906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3784a717540cSStefano Zampini } 3785a717540cSStefano Zampini 3786a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3787a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3788a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3789a717540cSStefano Zampini } 3790a717540cSStefano Zampini 3791a717540cSStefano Zampini 3792a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3793a717540cSStefano Zampini /* 3794a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3795a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3796a717540cSStefano Zampini 3797a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3798a717540cSStefano Zampini 3799a6b551f4SStefano 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) 3800a6b551f4SStefano Zampini 3801a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3802a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3803a717540cSStefano Zampini | ... | 3804a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3805a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3806a717540cSStefano Zampini 3807a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3808a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3809a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3810a6b551f4SStefano Zampini 3811a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3812a717540cSStefano Zampini */ 3813a717540cSStefano Zampini if (qr_needed) { 3814984c4197SStefano Zampini /* space to store Q */ 3815854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3816984c4197SStefano Zampini /* first we issue queries for optimal work */ 38173f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38183f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 38193f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3820984c4197SStefano Zampini lqr_work = -1; 38213f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3822984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3823984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3824785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3825984c4197SStefano Zampini lgqr_work = -1; 38263f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38273f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 38283f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 38293f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38303f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 38313f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3832984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3833984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3834785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3835984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3836785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3837984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3838785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3839a717540cSStefano Zampini /* allocating workspace for check */ 3840a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3841cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3842a717540cSStefano Zampini } 3843a717540cSStefano Zampini } 3844984c4197SStefano Zampini /* array to store whether a node is primal or not */ 38454641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3846473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 38470e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 384839e2fb2aSStefano Zampini if (i != total_primal_vertices) { 384939e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 38504641a718SStefano Zampini } 385139e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 385239e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 385339e2fb2aSStefano Zampini } 385439e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3855984c4197SStefano Zampini 3856a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 38579162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 38589162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 38594641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3860984c4197SStefano Zampini /* get constraint info */ 38619162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3862984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3863984c4197SStefano Zampini 3864984c4197SStefano Zampini if (pcbddc->dbg_flag) { 38659162d606SStefano 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); 3866674ae819SStefano Zampini } 3867984c4197SStefano Zampini 3868fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3869a717540cSStefano Zampini 3870a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3871a717540cSStefano Zampini if (pcbddc->dbg_flag) { 38729162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3873a717540cSStefano Zampini } 3874984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 38759162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3876984c4197SStefano Zampini 3877984c4197SStefano Zampini /* compute QR decomposition of constraints */ 38783f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 38793f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 38803f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3881674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 38823f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3883984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3884674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3885984c4197SStefano Zampini 3886984c4197SStefano Zampini /* explictly compute R^-T */ 3887984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3888984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 38893f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 38903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 38913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38923f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3893984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 38943f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3895984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3896984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3897984c4197SStefano Zampini 3898a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 38993f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39003f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39013f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39023f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3903984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39043f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3905984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3906984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3907984c4197SStefano Zampini 3908984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3909984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3910984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 39113f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39123f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39133f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 39163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3917984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39189162d606SStefano 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)); 3919984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39209162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3921984c4197SStefano Zampini 3922984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 39239162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3924984c4197SStefano Zampini /* insert cols for primal dofs */ 3925984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3926984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 39279162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3928906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3929984c4197SStefano Zampini } 3930984c4197SStefano Zampini /* insert cols for dual dofs */ 3931984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 39329162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3933984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 39349162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3935906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3936984c4197SStefano Zampini j++; 3937674ae819SStefano Zampini } 3938674ae819SStefano Zampini } 3939984c4197SStefano Zampini 3940984c4197SStefano Zampini /* check change of basis */ 3941984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3942984c4197SStefano Zampini PetscInt ii,jj; 3943984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3944c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3945c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3946c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3947c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3948c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3949c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3950984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3951cf5a6209SStefano 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)); 3952984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3953984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3954984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3955cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3956cf5a6209SStefano 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; 3957674ae819SStefano Zampini } 3958674ae819SStefano Zampini } 3959984c4197SStefano Zampini if (!valid_qr) { 396022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3961984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3962984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3963cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3964cf5a6209SStefano 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])); 3965674ae819SStefano Zampini } 3966cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3967cf5a6209SStefano 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])); 3968984c4197SStefano Zampini } 3969984c4197SStefano Zampini } 3970984c4197SStefano Zampini } 3971674ae819SStefano Zampini } else { 397222d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3973674ae819SStefano Zampini } 3974674ae819SStefano Zampini } 3975a717540cSStefano Zampini } else { /* simple transformation block */ 3976a717540cSStefano Zampini PetscInt row,col; 3977a6b551f4SStefano Zampini PetscScalar val,norm; 3978a6b551f4SStefano Zampini 3979a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39809162d606SStefano 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)); 3981a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 39829162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 39839162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3984bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 39859162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3986906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 39879162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3988a717540cSStefano Zampini } else { 3989a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 39909162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3991a717540cSStefano Zampini if (row != col) { 39929162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3993a717540cSStefano Zampini } else { 39949162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3995a717540cSStefano Zampini } 3996906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3997a717540cSStefano Zampini } 3998a717540cSStefano Zampini } 3999a717540cSStefano Zampini } 400098a51de6SStefano Zampini if (pcbddc->dbg_flag) { 400122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4002a717540cSStefano Zampini } 4003674ae819SStefano Zampini } 4004984c4197SStefano Zampini } else { 4005984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40069162d606SStefano 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); 4007674ae819SStefano Zampini } 4008674ae819SStefano Zampini } 4009674ae819SStefano Zampini } 4010a717540cSStefano Zampini 4011a717540cSStefano Zampini /* free workspace */ 4012a717540cSStefano Zampini if (qr_needed) { 4013984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4014cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4015984c4197SStefano Zampini } 4016984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4017984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4018984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4019984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4020984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4021674ae819SStefano Zampini } 4022a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4023906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4024906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4025906d46d4SStefano Zampini 4026906d46d4SStefano Zampini /* assembling of global change of variable */ 4027bbb9e6c6SStefano Zampini { 4028bbb9e6c6SStefano Zampini Mat tmat; 402916f15bc4SStefano Zampini PetscInt bs; 403016f15bc4SStefano Zampini 4031906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4032906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4033bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4034bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4035bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4036bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 403716f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 403816f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4039906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4040bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4041bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4042bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4043bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4044bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4045e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4046e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4047bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4048bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 4049906d46d4SStefano Zampini } 4050906d46d4SStefano Zampini /* check */ 4051906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4052906d46d4SStefano Zampini PetscReal error; 4053906d46d4SStefano Zampini Vec x,x_change; 4054906d46d4SStefano Zampini 4055906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4056906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4057906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4058906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4059e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4060e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4061bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4062e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4063e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4064906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4065906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4066906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4067906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4068bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4069906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4070906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4071906d46d4SStefano Zampini } 4072b96c3477SStefano Zampini 4073b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4074b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4075b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4076b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4077ac632422SStefano Zampini Mat S_new,tmat; 4078b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 4079b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 4080b087196eSStefano Zampini const PetscScalar *array; 4081b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 4082b087196eSStefano Zampini PetscInt i,n_V; 4083bbb9e6c6SStefano Zampini 4084bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 40856816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4086b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4087b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4088b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4089b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4090bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4091b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4092ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4093b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4094ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4095b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4096b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4097b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4098b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4099b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4100b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4101b087196eSStefano Zampini PetscScalar val; 4102b087196eSStefano Zampini PetscInt idx; 4103b087196eSStefano Zampini 4104b087196eSStefano Zampini idx = idxs_V[i]; 4105b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4106b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4107b087196eSStefano Zampini } 4108b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4109b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4110ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4111ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4112ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4113ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4114b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4115ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4116b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4117ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4118ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4119ac632422SStefano Zampini } 4120b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4121b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4122b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4123b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4124b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 4125b96c3477SStefano Zampini } 4126b96c3477SStefano Zampini } 4127906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 4128906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4129b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4130b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4131b9b85e73SStefano Zampini } 4132906d46d4SStefano Zampini 4133906d46d4SStefano Zampini /* set up change of basis context */ 4134906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4135906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4136906d46d4SStefano Zampini 4137906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4138906d46d4SStefano Zampini PetscInt global_size,local_size; 4139906d46d4SStefano Zampini 4140906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4141906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4142906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4143906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4144906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4145906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4146906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4147906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4148906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4149906d46d4SStefano Zampini } else { 4150906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4151906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4152906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4153906d46d4SStefano Zampini } 4154906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4155906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4156906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4157906d46d4SStefano Zampini } else { 4158906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4159906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4160906d46d4SStefano Zampini } 4161906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4162906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4163906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4164906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4165a8661815SStefano Zampini } else { 4166a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4167b9b85e73SStefano Zampini } 4168a717540cSStefano Zampini 41694f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 41704f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 41714f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 41724f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4173019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4174019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4175019a44ceSStefano Zampini pcbddc->local_primal_size++; 4176019a44ceSStefano Zampini } 4177019a44ceSStefano Zampini 4178019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4179727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4180727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 41819f47a83aSStefano 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); 4182c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 41830e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 41849f47a83aSStefano 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); 4185727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4186727cdba6SStefano Zampini } 41870e6343abSStefano Zampini } 41880e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4189727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4190727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4191727cdba6SStefano Zampini 4192a717540cSStefano Zampini /* flush dbg viewer */ 4193b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4194b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4195b8ffe317SStefano Zampini } 4196a717540cSStefano Zampini 4197e310c8b4SStefano Zampini /* free workspace */ 4198a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 41994641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 420008122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 42019162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 42029162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 420308122e43SStefano Zampini } else { 42049162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 42059162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 42069162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 420708122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 420808122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 42099162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 42109162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 421108122e43SStefano Zampini } 4212674ae819SStefano Zampini PetscFunctionReturn(0); 4213674ae819SStefano Zampini } 4214674ae819SStefano Zampini 4215674ae819SStefano Zampini #undef __FUNCT__ 4216674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4217674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4218674ae819SStefano Zampini { 4219674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4220674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4221674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 42227fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 4223674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4224674ae819SStefano Zampini 4225674ae819SStefano Zampini PetscFunctionBegin; 42268e61c736SStefano Zampini /* Reset previously computed graph */ 42278e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4228674ae819SStefano Zampini /* Init local Graph struct */ 42297fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 42303bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4231674ae819SStefano Zampini 4232575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 42335099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 42345099eff2SStefano 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); 4235575ad6abSStefano Zampini } 42369577ea80SStefano Zampini 4237674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 42384d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 42394d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 42404d379d7bSStefano Zampini PetscInt nvtxs; 4241e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4242674ae819SStefano Zampini 42434d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 42442fffb893SStefano Zampini 42452fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42462fffb893SStefano Zampini if (flg_row) { 42474d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4248b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 42492fffb893SStefano Zampini } 42502fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42519b28b941SStefano Zampini } else if (pcbddc->current_level && pcis->n_B) { /* just compute subdomain's connected components for coarser levels when the local boundary is not empty */ 42524d379d7bSStefano Zampini IS is_dummy; 42534d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 42544d379d7bSStefano Zampini PetscInt j,sum; 42554d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 42564d379d7bSStefano Zampini const PetscInt *idxs; 42574d379d7bSStefano Zampini PCBDDCGraph graph; 42584d379d7bSStefano Zampini PetscBT is_on_boundary; 42594d379d7bSStefano Zampini 42604d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 42614d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 42624d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 42634d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 42647fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 42654d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 4266e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4267e496cd5dSStefano Zampini if (flg_row) { 42684d379d7bSStefano Zampini graph->xadj = xadj; 42694d379d7bSStefano Zampini graph->adjncy = adjncy; 4270e496cd5dSStefano Zampini } 42714d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 42724d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4273e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42744d379d7bSStefano Zampini 42754d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 42769b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 42774d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 42784d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 42794d379d7bSStefano Zampini } 42804d379d7bSStefano Zampini } 42814d379d7bSStefano Zampini 4282e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 42834d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 42844d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 42854d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 42864d379d7bSStefano Zampini } 42874d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 42884d379d7bSStefano Zampini 4289e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 42904d379d7bSStefano Zampini sum = 0; 42914d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 42924d379d7bSStefano Zampini PetscInt sizecc = 0; 42934d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 42944d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 42954d379d7bSStefano Zampini sizecc++; 42964d379d7bSStefano Zampini } 42974d379d7bSStefano Zampini } 42984d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 42994d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 43004d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 43014d379d7bSStefano Zampini } 43024d379d7bSStefano Zampini } 43034d379d7bSStefano Zampini sum += sizecc*sizecc; 43044d379d7bSStefano Zampini } 43054d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 43064d379d7bSStefano Zampini sum = 0; 4307e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 43084d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 43094d379d7bSStefano Zampini cxadj[i] = sum; 43104d379d7bSStefano Zampini sum += temp; 43114d379d7bSStefano Zampini } 4312e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 43134d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 43144d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 43154d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 43164d379d7bSStefano Zampini PetscInt k,sizecc = 0; 43174d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 43184d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 43194d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 43204d379d7bSStefano Zampini sizecc++; 43214d379d7bSStefano Zampini } 43224d379d7bSStefano Zampini } 43234d379d7bSStefano Zampini } 43244d379d7bSStefano Zampini } 43254d379d7bSStefano Zampini } 43269b28b941SStefano Zampini if (sum) { 4327e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 43284d379d7bSStefano Zampini } else { 43294d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 43304d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 43314d379d7bSStefano Zampini } 43324d379d7bSStefano Zampini graph->xadj = 0; 43334d379d7bSStefano Zampini graph->adjncy = 0; 43344d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 43354d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 43364d379d7bSStefano Zampini } 4337674ae819SStefano Zampini } 43389b28b941SStefano Zampini if (pcbddc->dbg_flag) { 43399b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4340674ae819SStefano Zampini } 4341674ae819SStefano Zampini 434263602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 4343674ae819SStefano Zampini vertex_size = 1; 434463602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 434563602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 434695ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 434763602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4348e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 434963602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4350674ae819SStefano Zampini } 435163602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 435263602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 435363602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4354674ae819SStefano Zampini } 435563602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 4356674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 435763602bcaSStefano Zampini } else { 435863602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 435963602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4360854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 436163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 436263602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 436363602bcaSStefano Zampini } 436463602bcaSStefano Zampini } 4365674ae819SStefano Zampini } 4366674ae819SStefano Zampini 4367674ae819SStefano Zampini /* Setup of Graph */ 4368785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4369e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4370785d1243SStefano Zampini } 4371785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4372e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4373785d1243SStefano Zampini } 437430368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 437530368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 437630368db7SStefano Zampini } 437730368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4378674ae819SStefano Zampini 43794f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 43804f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 43814f1b2e48SStefano Zampini PetscInt *local_subs; 43824f1b2e48SStefano Zampini 43834f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 43844f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 43854f1b2e48SStefano Zampini const PetscInt *idxs; 43864f1b2e48SStefano Zampini PetscInt nl,j; 43874f1b2e48SStefano Zampini 43884f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 43894f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43904f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 43914f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 43924f1b2e48SStefano Zampini } 43934f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43944f1b2e48SStefano Zampini } 43954f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 43964f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 43974f1b2e48SStefano Zampini } 43984f1b2e48SStefano Zampini 4399674ae819SStefano Zampini /* Graph's connected components analysis */ 4400674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4401674ae819SStefano Zampini 4402674ae819SStefano Zampini /* print some info to stdout */ 4403674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4404302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4405674ae819SStefano Zampini } 4406fb180af4SStefano Zampini 4407fb180af4SStefano Zampini /* mark topography has done */ 4408fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 4409674ae819SStefano Zampini PetscFunctionReturn(0); 4410674ae819SStefano Zampini } 4411674ae819SStefano Zampini 4412dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4413674ae819SStefano Zampini #undef __FUNCT__ 4414674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4415dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4416674ae819SStefano Zampini { 4417dc456d91SStefano Zampini PetscSF sf; 4418dc456d91SStefano Zampini PetscLayout map; 4419dc456d91SStefano Zampini const PetscInt *idxs; 4420dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4421dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4422dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4423dc456d91SStefano Zampini PetscMPIInt commsize; 4424674ae819SStefano Zampini PetscBool first_found; 4425674ae819SStefano Zampini PetscErrorCode ierr; 4426674ae819SStefano Zampini 4427674ae819SStefano Zampini PetscFunctionBegin; 4428dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4429dc456d91SStefano Zampini if (subset_mult) { 4430dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4431dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4432dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4433674ae819SStefano Zampini } 4434dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4435dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4436dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4437dc456d91SStefano Zampini for (i=0;i<n;i++) { 4438dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4439dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4440674ae819SStefano Zampini } 4441dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4442dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4443dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4444dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4445dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4446dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4447dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4448dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4449dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4450dc456d91SStefano Zampini 4451dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4452dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4453dc456d91SStefano Zampini if (subset_mult) { 4454dc456d91SStefano Zampini const PetscInt* idxs_mult; 4455dc456d91SStefano Zampini 4456dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4457dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4458dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4459674ae819SStefano Zampini } else { 4460dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4461674ae819SStefano Zampini } 4462dc456d91SStefano Zampini /* local size of new subset */ 4463dc456d91SStefano Zampini n_n = 0; 4464dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4465dc456d91SStefano Zampini 4466dc456d91SStefano Zampini /* global indexes in layout */ 4467dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4468dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4469dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4470dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4471dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4472dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4473dc456d91SStefano Zampini 4474dc456d91SStefano Zampini /* reduce from leaves to roots */ 4475dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 447664a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 447764a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4478dc456d91SStefano Zampini 4479dc456d91SStefano Zampini /* count indexes in local part of layout */ 4480674ae819SStefano Zampini nlocals = 0; 4481674ae819SStefano Zampini first_index = -1; 4482674ae819SStefano Zampini first_found = PETSC_FALSE; 4483dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4484dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4485674ae819SStefano Zampini first_found = PETSC_TRUE; 4486674ae819SStefano Zampini first_index = i; 4487674ae819SStefano Zampini } 4488dc456d91SStefano Zampini nlocals += root_data[i]; 4489674ae819SStefano Zampini } 4490dc456d91SStefano Zampini 4491dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 44925fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4493dc456d91SStefano Zampini start = 0; 449464a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44955fa240b1SStefano Zampini #else 449664a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44975fa240b1SStefano Zampini start = start-nlocals; 44985fa240b1SStefano Zampini #endif 44995fa240b1SStefano Zampini 4500dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4501dc456d91SStefano Zampini *N_n = start + nlocals; 4502dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4503dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4504674ae819SStefano Zampini } 45055fa240b1SStefano Zampini 45065fa240b1SStefano Zampini /* adapt root data with cumulative */ 4507674ae819SStefano Zampini if (first_found) { 4508dc456d91SStefano Zampini PetscInt old_index; 4509dc456d91SStefano Zampini 4510dc456d91SStefano Zampini root_data[first_index] += start; 4511674ae819SStefano Zampini old_index = first_index; 4512dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4513dc456d91SStefano Zampini if (root_data[i]) { 4514dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4515674ae819SStefano Zampini old_index = i; 4516674ae819SStefano Zampini } 4517674ae819SStefano Zampini } 4518674ae819SStefano Zampini } 4519dc456d91SStefano Zampini 4520dc456d91SStefano Zampini /* from roots to leaves */ 4521dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4522dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4523dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4524dc456d91SStefano Zampini 4525dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4526dc456d91SStefano Zampini if (subset_mult) { 4527dc456d91SStefano Zampini const PetscInt* idxs_mult; 4528dc456d91SStefano Zampini PetscInt cum; 4529dc456d91SStefano Zampini 4530dc456d91SStefano Zampini cum = 0; 4531dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4532dc456d91SStefano Zampini for (i=0;i<n;i++) { 4533dc456d91SStefano Zampini PetscInt j; 4534dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4535674ae819SStefano Zampini } 4536dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4537674ae819SStefano Zampini } else { 4538dc456d91SStefano Zampini for (i=0;i<n;i++) { 4539dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4540674ae819SStefano Zampini } 4541674ae819SStefano Zampini } 4542dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4543dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4544674ae819SStefano Zampini PetscFunctionReturn(0); 4545674ae819SStefano Zampini } 45469a7d3425SStefano Zampini 45479a7d3425SStefano Zampini #undef __FUNCT__ 45489a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 45499a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 45509a7d3425SStefano Zampini { 45519a7d3425SStefano Zampini PetscInt i,j; 45529a7d3425SStefano Zampini PetscScalar *alphas; 45539a7d3425SStefano Zampini PetscErrorCode ierr; 45549a7d3425SStefano Zampini 45559a7d3425SStefano Zampini PetscFunctionBegin; 45569a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4557785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 45589a7d3425SStefano Zampini for (i=0;i<n;i++) { 45599a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 45609a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 45619a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 45629a7d3425SStefano Zampini } 45639a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 45649a7d3425SStefano Zampini PetscFunctionReturn(0); 45659a7d3425SStefano Zampini } 45669a7d3425SStefano Zampini 4567e7931f94SStefano Zampini #undef __FUNCT__ 456870cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4569b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4570e7931f94SStefano Zampini { 457152e5ac9dSStefano Zampini IS ranks_send_to; 4572e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4573e7931f94SStefano Zampini PetscMPIInt size,rank,color; 457452e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 457552e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 45763837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 45772b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4578e7931f94SStefano Zampini PetscSubcomm subcomm; 457952e5ac9dSStefano Zampini PetscErrorCode ierr; 4580a57a6d2fSStefano Zampini 4581e7931f94SStefano Zampini PetscFunctionBegin; 45822b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 45832b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 45842b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4585e7931f94SStefano Zampini 4586e7931f94SStefano Zampini /* Get info on mapping */ 45873bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 45883bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4589e7931f94SStefano Zampini 4590e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4591785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4592e7931f94SStefano Zampini xadj[0] = 0; 4593e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4594785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4595785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4596e7931f94SStefano Zampini 45972b510759SStefano Zampini if (threshold) { 4598d023bfaeSStefano Zampini PetscInt xadj_count = 0; 45992b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4600d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4601d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4602d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4603d023bfaeSStefano Zampini xadj_count++; 4604e7931f94SStefano Zampini } 4605e7931f94SStefano Zampini } 4606d023bfaeSStefano Zampini xadj[1] = xadj_count; 4607c8587f34SStefano Zampini } else { 4608e7931f94SStefano Zampini if (xadj[1]) { 4609e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4610e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4611c8587f34SStefano Zampini } 4612e7931f94SStefano Zampini } 46133bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4614e7931f94SStefano Zampini if (use_square) { 4615e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4616e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4617e7931f94SStefano Zampini } 4618e7931f94SStefano Zampini } 4619e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4620e7931f94SStefano Zampini 46213837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4622e7931f94SStefano Zampini 4623e7931f94SStefano Zampini /* 4624e7931f94SStefano Zampini Restrict work on active processes only. 4625e7931f94SStefano Zampini */ 4626e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4627e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4628e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 46292b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4630d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4631e7931f94SStefano Zampini if (color) { 4632e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4633e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4634e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4635c8587f34SStefano Zampini } else { 463652e5ac9dSStefano Zampini Mat subdomain_adj; 463752e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 463852e5ac9dSStefano Zampini MatPartitioning partitioner; 463952e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 464052e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4641b0c7d250SStefano Zampini PetscBool aggregate; 4642b0c7d250SStefano Zampini 4643306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4644785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4645e7931f94SStefano Zampini prank = rank; 4646306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 46478002ef2cSStefano Zampini /* 4648e7931f94SStefano Zampini for (i=0;i<size;i++) { 4649e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4650c8587f34SStefano Zampini } 46518002ef2cSStefano Zampini */ 4652e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4653e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4654c8587f34SStefano Zampini } 4655e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4656b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4657b0c7d250SStefano Zampini if (aggregate) { 4658b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4659b0c7d250SStefano Zampini PetscMPIInt nrank; 4660b0c7d250SStefano Zampini PetscScalar *vals; 4661b0c7d250SStefano Zampini 4662b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4663b0c7d250SStefano Zampini lrows = 0; 4664b0c7d250SStefano Zampini if (nrank<redprocs) { 4665b0c7d250SStefano Zampini lrows = size/redprocs; 4666b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4667b0c7d250SStefano Zampini } 46685fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4669b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4670b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4671b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4672b0c7d250SStefano Zampini row = nrank; 4673b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4674b0c7d250SStefano Zampini cols = adjncy; 4675b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4676b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4677b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4678b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4679b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 468052e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 468152e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 468252e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4683b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4684b0c7d250SStefano Zampini } else { 4685306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4686b0c7d250SStefano Zampini } 468722b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4688e7931f94SStefano Zampini 4689e7931f94SStefano Zampini /* Partition */ 4690306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4691e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4692e7931f94SStefano Zampini if (use_vwgt) { 46933837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4694e7931f94SStefano Zampini v_wgt[0] = local_size; 4695e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4696c8587f34SStefano Zampini } 469728143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 469828143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4699e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4700e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 470122b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4702e7931f94SStefano Zampini 470352e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 470452e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 470552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 470652e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4707b0c7d250SStefano Zampini if (!redprocs) { 4708b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 470928143c3dSStefano Zampini } else { 4710b0c7d250SStefano Zampini PetscInt idxs[1]; 4711b0c7d250SStefano Zampini PetscMPIInt tag; 4712b0c7d250SStefano Zampini MPI_Request *reqs; 4713b0c7d250SStefano Zampini 4714b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4715b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4716b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4717b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 471828143c3dSStefano Zampini } 4719b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4720b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4721b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4722b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4723e7931f94SStefano Zampini } 472452e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4725e7931f94SStefano Zampini /* clean up */ 4726e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 472752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4728e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4729e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4730e7931f94SStefano Zampini } 4731e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4732e7931f94SStefano Zampini 4733e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4734e7931f94SStefano Zampini i = 1; 4735e7931f94SStefano Zampini if (color) i=0; 4736e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4737e7931f94SStefano Zampini /* get back IS */ 4738e7931f94SStefano Zampini *is_sends = ranks_send_to; 4739e7931f94SStefano Zampini PetscFunctionReturn(0); 4740e7931f94SStefano Zampini } 4741e7931f94SStefano Zampini 4742e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4743e7931f94SStefano Zampini 4744e7931f94SStefano Zampini #undef __FUNCT__ 4745e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 474653a05cb3SStefano 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[]) 4747e7931f94SStefano Zampini { 474870cf5478SStefano Zampini Mat local_mat; 4749e7931f94SStefano Zampini IS is_sends_internal; 47509d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 475128143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 47529d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4753e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4754e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4755e7931f94SStefano Zampini const PetscInt* is_indices; 4756e7931f94SStefano Zampini MatType new_local_type; 4757e7931f94SStefano Zampini /* buffers */ 4758e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 475928143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 47609d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4761e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4762e7931f94SStefano Zampini /* MPI */ 476328143c3dSStefano Zampini MPI_Comm comm,comm_n; 476428143c3dSStefano Zampini PetscSubcomm subcomm; 4765e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 476628143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 476728143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 476828143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 476928143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 477028143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4771e7931f94SStefano Zampini PetscErrorCode ierr; 4772e7931f94SStefano Zampini 4773e7931f94SStefano Zampini PetscFunctionBegin; 477428143c3dSStefano Zampini /* TODO: add missing checks */ 477528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 477628143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 477728143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 477828143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4779e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 478028143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4781e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4782e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4783e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4784e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4785e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 478628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 478770cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 478870cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 478928143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 479070cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 479170cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 479270cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 479370cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 479470cf5478SStefano Zampini } 4795e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4796e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4797e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4798e7931f94SStefano Zampini if (!is_sends) { 479928143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4800b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4801c8587f34SStefano Zampini } else { 4802e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4803e7931f94SStefano Zampini is_sends_internal = is_sends; 4804c8587f34SStefano Zampini } 4805e7931f94SStefano Zampini 4806e7931f94SStefano Zampini /* get comm */ 4807a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4808e7931f94SStefano Zampini 4809e7931f94SStefano Zampini /* compute number of sends */ 4810e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4811e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4812e7931f94SStefano Zampini 4813e7931f94SStefano Zampini /* compute number of receives */ 4814e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4815785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4816e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4817e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4818e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4819e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4820e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4821e7931f94SStefano Zampini 482228143c3dSStefano Zampini /* restrict comm if requested */ 482328143c3dSStefano Zampini subcomm = 0; 482428143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 482528143c3dSStefano Zampini if (restrict_comm) { 4826779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4827779c1cceSStefano Zampini 482828143c3dSStefano Zampini color = 0; 482953a05cb3SStefano Zampini if (restrict_full) { 483053a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 483153a05cb3SStefano Zampini } else { 483253a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 483353a05cb3SStefano Zampini } 483428143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 483528143c3dSStefano Zampini subcommsize = commsize - subcommsize; 483628143c3dSStefano Zampini /* check if reuse has been requested */ 483728143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 483828143c3dSStefano Zampini if (*mat_n) { 483928143c3dSStefano Zampini PetscMPIInt subcommsize2; 484028143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 484128143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 484228143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 484328143c3dSStefano Zampini } else { 484428143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 484528143c3dSStefano Zampini } 484628143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4847779c1cceSStefano Zampini PetscMPIInt rank; 4848779c1cceSStefano Zampini 4849779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 485028143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 485128143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 485228143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4853306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 485428143c3dSStefano Zampini } 485528143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 485628143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 485728143c3dSStefano Zampini } else { 485828143c3dSStefano Zampini comm_n = comm; 485928143c3dSStefano Zampini } 486028143c3dSStefano Zampini 4861e7931f94SStefano Zampini /* prepare send/receive buffers */ 4862785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4863e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4864785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4865e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 486628143c3dSStefano Zampini if (nis) { 4867854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 486828143c3dSStefano Zampini } 4869e7931f94SStefano Zampini 487028143c3dSStefano Zampini /* Get data from local matrices */ 4871e7931f94SStefano Zampini if (!isdense) { 4872a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4873e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4874e7931f94SStefano Zampini /* 4875e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4876e7931f94SStefano Zampini send_buffer_idxs should contain: 4877e7931f94SStefano Zampini - MatType_PRIVATE type 4878e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4879e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4880e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4881e7931f94SStefano Zampini */ 4882e7931f94SStefano Zampini } else { 4883e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 48843bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4885854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4886e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4887e7931f94SStefano Zampini send_buffer_idxs[1] = i; 48883bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4889e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 48903bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4891e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4892e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4893e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4894e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4895c8587f34SStefano Zampini } 4896c8587f34SStefano Zampini } 4897e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 489828143c3dSStefano Zampini /* additional is (if any) */ 489928143c3dSStefano Zampini if (nis) { 490028143c3dSStefano Zampini PetscMPIInt psum; 490128143c3dSStefano Zampini PetscInt j; 490228143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 490328143c3dSStefano Zampini PetscInt plen; 490428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 490528143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 490628143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 490728143c3dSStefano Zampini } 4908854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 490928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 491028143c3dSStefano Zampini PetscInt plen; 491128143c3dSStefano Zampini const PetscInt *is_array_idxs; 491228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 491328143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 491428143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 491528143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 491628143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 491728143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 491828143c3dSStefano Zampini } 491928143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 492028143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 492128143c3dSStefano Zampini } 492228143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 492328143c3dSStefano Zampini } 492428143c3dSStefano Zampini 4925e7931f94SStefano Zampini buf_size_idxs = 0; 4926e7931f94SStefano Zampini buf_size_vals = 0; 492728143c3dSStefano Zampini buf_size_idxs_is = 0; 4928e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4929e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4930e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 493128143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4932e7931f94SStefano Zampini } 4933785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4934785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 493595ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4936e7931f94SStefano Zampini 4937e7931f94SStefano Zampini /* get new tags for clean communications */ 4938e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4939e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 494028143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4941e7931f94SStefano Zampini 4942e7931f94SStefano Zampini /* allocate for requests */ 4943785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4944785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 494595ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4946785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4947785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 494895ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4949e7931f94SStefano Zampini 4950e7931f94SStefano Zampini /* communications */ 4951e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4952e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 495328143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4954e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4955e7931f94SStefano Zampini source_dest = onodes[i]; 4956e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4957e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4958e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4959e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 496028143c3dSStefano Zampini if (nis) { 496128143c3dSStefano 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); 496228143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 496328143c3dSStefano Zampini } 4964e7931f94SStefano Zampini } 4965e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4966e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4967e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4968e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 496928143c3dSStefano Zampini if (nis) { 497028143c3dSStefano 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); 497128143c3dSStefano Zampini } 4972e7931f94SStefano Zampini } 4973e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4974e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4975e7931f94SStefano Zampini 4976e7931f94SStefano Zampini /* assemble new l2g map */ 4977e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4978e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49799d30be91SStefano Zampini new_local_rows = 0; 4980e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49819d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4982e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4983e7931f94SStefano Zampini } 49849d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4985e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49869d30be91SStefano Zampini new_local_rows = 0; 4987e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49889d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 49899d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4990e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4991e7931f94SStefano Zampini } 49929d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 49939d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4994e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4995e7931f94SStefano Zampini 4996e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4997e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4998e7931f94SStefano 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) */ 4999e7931f94SStefano Zampini if (n_recvs) { 500028143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5001e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5002e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5003e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5004e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5005e7931f94SStefano Zampini break; 5006e7931f94SStefano Zampini } 5007e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5008e7931f94SStefano Zampini } 5009e7931f94SStefano Zampini switch (new_local_type_private) { 501028143c3dSStefano Zampini case MATDENSE_PRIVATE: 501128143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5012e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5013e7931f94SStefano Zampini bs = 1; 501428143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 501528143c3dSStefano Zampini new_local_type = MATSEQDENSE; 501628143c3dSStefano Zampini bs = 1; 501728143c3dSStefano Zampini } 5018e7931f94SStefano Zampini break; 5019e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5020e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5021e7931f94SStefano Zampini bs = 1; 5022e7931f94SStefano Zampini break; 5023e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5024e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5025e7931f94SStefano Zampini break; 5026e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5027e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5028e7931f94SStefano Zampini break; 5029e7931f94SStefano Zampini default: 50309d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5031e7931f94SStefano Zampini break; 5032e7931f94SStefano Zampini } 503328143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 503428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 503528143c3dSStefano Zampini bs = 1; 5036e7931f94SStefano Zampini } 5037e7931f94SStefano Zampini 503870cf5478SStefano Zampini /* create MATIS object if needed */ 503970cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5040e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5041e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 504270cf5478SStefano Zampini } else { 504370cf5478SStefano Zampini /* it also destroys the local matrices */ 504470cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 504570cf5478SStefano Zampini } 504670cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5047e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 50489d30be91SStefano Zampini 50499d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 50509d30be91SStefano Zampini 50519d30be91SStefano Zampini /* Global to local map of received indices */ 50529d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 50539d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 50549d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 50559d30be91SStefano Zampini 50569d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 50579d30be91SStefano Zampini buf_size_idxs = 0; 50589d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50599d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 50609d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 50619d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 50629d30be91SStefano Zampini } 50639d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 50649d30be91SStefano Zampini 50659d30be91SStefano Zampini /* set preallocation */ 50669d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 50679d30be91SStefano Zampini if (!newisdense) { 50689d30be91SStefano Zampini PetscInt *new_local_nnz=0; 50699d30be91SStefano Zampini 50709d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 50719d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 50729d30be91SStefano Zampini if (n_recvs) { 50739d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 50749d30be91SStefano Zampini } 50759d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50769d30be91SStefano Zampini PetscInt j; 50779d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 50789d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 50799d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 50809d30be91SStefano Zampini } 50819d30be91SStefano Zampini } else { 50829d30be91SStefano Zampini /* TODO */ 50839d30be91SStefano Zampini } 50849d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 50859d30be91SStefano Zampini } 50869d30be91SStefano Zampini if (new_local_nnz) { 50879d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 50889d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 50899d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 50909d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50919d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 50929d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50939d30be91SStefano Zampini } else { 50949d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50959d30be91SStefano Zampini } 50969d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 50979d30be91SStefano Zampini } else { 50989d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50999d30be91SStefano Zampini } 5100e7931f94SStefano Zampini 5101e7931f94SStefano Zampini /* set values */ 5102e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 51039d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5104e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5105e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5106e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 51079d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5108e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5109e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5110e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 511128143c3dSStefano Zampini } else { 511228143c3dSStefano Zampini /* TODO */ 5113e7931f94SStefano Zampini } 5114e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5115e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5116e7931f94SStefano Zampini } 5117e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5118e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 511970cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 512070cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 51219d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 51229d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5123e7931f94SStefano Zampini 5124dfd14d43SStefano Zampini #if 0 512528143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5126e7931f94SStefano Zampini Vec lvec,rvec; 5127e7931f94SStefano Zampini PetscReal infty_error; 5128e7931f94SStefano Zampini 51292a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5130e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5131e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5132e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 513370cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5134e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5135e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5136e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5137e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5138e7931f94SStefano Zampini } 513928143c3dSStefano Zampini #endif 5140e7931f94SStefano Zampini 514128143c3dSStefano Zampini /* assemble new additional is (if any) */ 514228143c3dSStefano Zampini if (nis) { 514328143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 514428143c3dSStefano Zampini 514528143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5146854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 514728143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 514828143c3dSStefano Zampini psum = 0; 514928143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 515028143c3dSStefano Zampini for (j=0;j<nis;j++) { 515128143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 515228143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 515328143c3dSStefano Zampini psum += plen; 515428143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 515528143c3dSStefano Zampini } 515628143c3dSStefano Zampini } 5157854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5158854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 515928143c3dSStefano Zampini for (i=1;i<nis;i++) { 516028143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 516128143c3dSStefano Zampini } 516228143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 516328143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 516428143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 516528143c3dSStefano Zampini for (j=0;j<nis;j++) { 516628143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 516728143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 516828143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 516928143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 517028143c3dSStefano Zampini } 517128143c3dSStefano Zampini } 517228143c3dSStefano Zampini for (i=0;i<nis;i++) { 517328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 517428143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 517528143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 517628143c3dSStefano Zampini } 517728143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 517828143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 517928143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 518028143c3dSStefano Zampini } 5181e7931f94SStefano Zampini /* free workspace */ 518228143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5183e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5184e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5185e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5186e7931f94SStefano Zampini if (isdense) { 5187e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5188e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5189e7931f94SStefano Zampini } else { 5190e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5191e7931f94SStefano Zampini } 519228143c3dSStefano Zampini if (nis) { 519328143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 519428143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 519528143c3dSStefano Zampini } 5196e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5197e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 519828143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5199e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5200e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 520128143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5202e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5203e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5204e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5205e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5206e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 520728143c3dSStefano Zampini if (nis) { 520828143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 520928143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 521028143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 521128143c3dSStefano Zampini } 521228143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 521328143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 521428143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 521528143c3dSStefano Zampini for (i=0;i<nis;i++) { 521628143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 521728143c3dSStefano Zampini } 521853a05cb3SStefano Zampini *mat_n = NULL; 521928143c3dSStefano Zampini } 5220e7931f94SStefano Zampini PetscFunctionReturn(0); 5221e7931f94SStefano Zampini } 5222a57a6d2fSStefano Zampini 522312edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5224af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 522512edc857SStefano Zampini 5226c8587f34SStefano Zampini #undef __FUNCT__ 5227c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5228c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5229c8587f34SStefano Zampini { 5230c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5231c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 523220a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 52339881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 523420a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 52356e683305SStefano Zampini IS coarse_is,*isarray; 52366e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 523730368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5238f9eb5b7dSStefano Zampini PC pc_temp; 5239c8587f34SStefano Zampini PCType coarse_pc_type; 5240c8587f34SStefano Zampini KSPType coarse_ksp_type; 5241f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 52424f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 52436e683305SStefano Zampini Mat t_coarse_mat_is; 52446e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 52456e683305SStefano Zampini PetscMPIInt all_procs; 524674e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 524768457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 524822bc73bbSStefano Zampini PetscScalar *array; 52499881197aSStefano Zampini PetscErrorCode ierr; 5250fdc09c96SStefano Zampini 5251c8587f34SStefano Zampini PetscFunctionBegin; 5252c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 525368457ee5SStefano 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 */ 5254fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 52555a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5256fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5257f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5258f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5259f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5260fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 526151bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 526251bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5263dc4bcba2SStefano Zampini PC pc; 5264dc4bcba2SStefano Zampini PetscBool isbddc; 5265dc4bcba2SStefano Zampini 5266dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5267dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5268dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5269dc4bcba2SStefano Zampini if (isbddc) { 5270dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 5271dc4bcba2SStefano Zampini } 5272727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 5273fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5274fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5275fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5276f4ddd8eeSStefano Zampini } 5277fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5278fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5279f4ddd8eeSStefano Zampini } 528070cf5478SStefano Zampini /* reset any subassembling information */ 528170cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 52826e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 52836e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5284fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5285f4ddd8eeSStefano Zampini } 5286c8587f34SStefano Zampini 52876e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 52882b510759SStefano Zampini im_active = !!(pcis->n); 52892b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 52906e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 52916e683305SStefano Zampini void_procs = all_procs-active_procs; 52926e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 529374e2c79eSStefano Zampini redist = PETSC_FALSE; 529422bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 52956e683305SStefano Zampini csin_ml = PETSC_TRUE; 52966e683305SStefano Zampini ncoarse_ml = void_procs; 5297779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5298779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 52996e683305SStefano Zampini csin_ds = PETSC_TRUE; 530018a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 530118a45a71SStefano Zampini redist = PETSC_TRUE; 530218a45a71SStefano Zampini } else { 53036e683305SStefano Zampini csin_ds = PETSC_TRUE; 5304779c1cceSStefano Zampini ncoarse_ds = active_procs; 5305779c1cceSStefano Zampini redist = PETSC_TRUE; 530618a45a71SStefano Zampini } 53076e683305SStefano Zampini } else { 53086e683305SStefano Zampini csin_ml = PETSC_FALSE; 53096e683305SStefano Zampini ncoarse_ml = all_procs; 53106e683305SStefano Zampini if (void_procs) { 53116e683305SStefano Zampini csin_ds = PETSC_TRUE; 53126e683305SStefano Zampini ncoarse_ds = void_procs; 53136e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 53146e683305SStefano Zampini } else { 5315779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 531674e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 531774e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 531874e2c79eSStefano Zampini redist = PETSC_TRUE; 531974e2c79eSStefano Zampini } else { 53206e683305SStefano Zampini csin_ds = PETSC_FALSE; 53216e683305SStefano Zampini ncoarse_ds = all_procs; 53226e683305SStefano Zampini } 53236e683305SStefano Zampini } 532474e2c79eSStefano Zampini } 53256e683305SStefano Zampini 53266e683305SStefano Zampini /* 53276e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 53286e683305SStefano Zampini - we have not exceeded the number of levels requested 53296e683305SStefano Zampini - we can actually subassemble the active processes 53306e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 53316e683305SStefano Zampini */ 53326e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 53336e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 53346e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 53356e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 53366e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5337f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 53382b510759SStefano Zampini } else { 5339f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5340c8587f34SStefano Zampini } 5341c8587f34SStefano Zampini } 53426e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 53436e683305SStefano Zampini if (multilevel_allowed) { 53446e683305SStefano Zampini ncoarse = ncoarse_ml; 53456e683305SStefano Zampini csin = csin_ml; 534658da7f69SStefano Zampini redist = PETSC_FALSE; 53476e683305SStefano Zampini } else { 53486e683305SStefano Zampini ncoarse = ncoarse_ds; 53496e683305SStefano Zampini csin = csin_ds; 53506e683305SStefano Zampini } 5351e7931f94SStefano Zampini 5352abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5353abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5354abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5355abbbba34SStefano Zampini 5356abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 535722bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 535822bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 535922bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 536022bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5361e176bc59SStefano 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); 53626e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 53636e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53646e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5365abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5366abbbba34SStefano Zampini 53676e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 536830368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 53696e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 53706e683305SStefano Zampini const PetscInt *idxs; 53716e683305SStefano Zampini ISLocalToGlobalMapping tmap; 53726e683305SStefano Zampini 53736e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 53740be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 53756e683305SStefano Zampini /* allocate space for temporary storage */ 5376854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5377854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 53786e683305SStefano Zampini /* allocate for IS array */ 53796e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 53806e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 538130368db7SStefano Zampini nisvert = 0; 538230368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 538330368db7SStefano Zampini nisvert = 1; 538430368db7SStefano Zampini } 538530368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5386854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 53876e683305SStefano Zampini /* dofs splitting */ 53886e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 53896e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 53906e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 53916e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53926e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53936e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53946e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 539530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 53966e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 53976e683305SStefano Zampini } 53986e683305SStefano Zampini /* neumann boundaries */ 53996e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 54006e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 54016e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 54026e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54036e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54046e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54056e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 540630368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 54076e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 54086e683305SStefano Zampini } 540930368db7SStefano Zampini /* primal vertices (benign) */ 541030368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 541130368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 541230368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 541330368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 541430368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 541530368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 541630368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 541730368db7SStefano Zampini } 54186e683305SStefano Zampini /* free memory */ 54196e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 54206e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 54216e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 54226e683305SStefano Zampini } else { 54236e683305SStefano Zampini nis = 0; 54246e683305SStefano Zampini nisdofs = 0; 54256e683305SStefano Zampini nisneu = 0; 542630368db7SStefano Zampini nisvert = 0; 54276e683305SStefano Zampini isarray = NULL; 54286e683305SStefano Zampini } 54296e683305SStefano Zampini /* destroy no longer needed map */ 54306e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 54316e683305SStefano Zampini 54326e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 54336e683305SStefano Zampini coarse_mat_is = NULL; 54346e683305SStefano Zampini if (csin) { 54356e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 543674e2c79eSStefano Zampini if (redist) { 543774e2c79eSStefano Zampini PetscMPIInt rank; 5438779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 543974e2c79eSStefano Zampini 544074e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 544158da7f69SStefano Zampini spc = active_procs/ncoarse; 544258da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5443779c1cceSStefano Zampini if (im_active) { 5444779c1cceSStefano Zampini destsize = 1; 544574e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 544674e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 544774e2c79eSStefano Zampini } else { 544874e2c79eSStefano Zampini dest[0] = rank/(spc+1); 544974e2c79eSStefano Zampini } 545074e2c79eSStefano Zampini } else { 5451779c1cceSStefano Zampini destsize = 0; 54526e683305SStefano Zampini } 5453779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5454779c1cceSStefano Zampini } else if (csin_type_simple) { 54556e683305SStefano Zampini PetscMPIInt rank; 54566e683305SStefano Zampini PetscInt issize,isidx; 5457779c1cceSStefano Zampini 54586e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 54596e683305SStefano Zampini if (im_active) { 54606e683305SStefano Zampini issize = 1; 54616e683305SStefano Zampini isidx = (PetscInt)rank; 54626e683305SStefano Zampini } else { 54636e683305SStefano Zampini issize = 0; 54646e683305SStefano Zampini isidx = -1; 54656e683305SStefano Zampini } 54666e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5467779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5468b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54696e683305SStefano Zampini } 5470779c1cceSStefano Zampini 5471779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5472779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5473779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5474779c1cceSStefano Zampini PetscInt *coarse_candidates; 5475779c1cceSStefano Zampini const PetscInt* tisindices; 5476779c1cceSStefano Zampini 5477779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5478779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5479779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5480779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5481779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5482779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5483779c1cceSStefano Zampini } 5484779c1cceSStefano Zampini } 5485779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5486779c1cceSStefano Zampini 5487779c1cceSStefano Zampini 54886e683305SStefano Zampini if (pcbddc->dbg_flag) { 54896e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54906e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 54916e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54926e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5493779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 54946e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 54956e683305SStefano Zampini } 54966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 54976e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54986e683305SStefano Zampini } 54996e683305SStefano Zampini /* shift the pattern on coarse candidates */ 55006e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 55016e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5502854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 55036e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 55046e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 55056e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 55066e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 55076e683305SStefano Zampini } 55086e683305SStefano Zampini if (pcbddc->dbg_flag) { 55096e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55106e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 55116e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55126e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55136e683305SStefano Zampini } 5514779c1cceSStefano Zampini } 55156e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 551653a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 551753a05cb3SStefano 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); 551853a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 551953a05cb3SStefano 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); 552053a05cb3SStefano Zampini } 55216e683305SStefano Zampini } else { 55226e683305SStefano Zampini if (pcbddc->dbg_flag) { 55236e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 55256e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55266e683305SStefano Zampini } 55276e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 55286e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 55296e683305SStefano Zampini } 55306e683305SStefano Zampini 55316e683305SStefano Zampini /* create local to global scatters for coarse problem */ 553268457ee5SStefano Zampini if (compute_vecs) { 55336e683305SStefano Zampini PetscInt lrows; 55346e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 55356e683305SStefano Zampini if (coarse_mat_is) { 55366e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 55376e683305SStefano Zampini } else { 55386e683305SStefano Zampini lrows = 0; 55396e683305SStefano Zampini } 55406e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 55416e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 55426e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 55436e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55446e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55456e683305SStefano Zampini } 55466e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 55476e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5548c8587f34SStefano Zampini 5549f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5550f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5551f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5552f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5553f9eb5b7dSStefano Zampini } else { 5554f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5555f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5556c8587f34SStefano Zampini } 5557c8587f34SStefano Zampini 55586e683305SStefano Zampini /* print some info if requested */ 55596e683305SStefano Zampini if (pcbddc->dbg_flag) { 55606e683305SStefano Zampini if (!multilevel_allowed) { 55616e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55626e683305SStefano Zampini if (multilevel_requested) { 55636e683305SStefano 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); 55646e683305SStefano Zampini } else if (pcbddc->max_levels) { 55656e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 55666e683305SStefano Zampini } 55676e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55686e683305SStefano Zampini } 55696e683305SStefano Zampini } 55706e683305SStefano Zampini 5571f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 55726e683305SStefano Zampini if (coarse_mat_is) { 55736e683305SStefano Zampini MatReuse coarse_mat_reuse; 55746a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 55756e683305SStefano Zampini if (pcbddc->dbg_flag) { 55766e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 55776e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 55786e683305SStefano Zampini } 5579f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5580312be037SStefano Zampini char prefix[256],str_level[16]; 5581e604994aSStefano Zampini size_t len; 55826e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5583422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5584c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5585f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 55865f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5587c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 55886e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5589c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5590c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5591e604994aSStefano Zampini /* prefix */ 5592e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5593e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5594e604994aSStefano Zampini if (!pcbddc->current_level) { 5595e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5596e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5597c8587f34SStefano Zampini } else { 5598e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5599312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5600312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 560134d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5602312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5603e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5604e604994aSStefano Zampini } 5605e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 56063e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 56073e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 56083e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 56093e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5610f9eb5b7dSStefano Zampini /* allow user customization */ 5611f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 56123e3c6dadSStefano Zampini } 56133e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 561451bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 56153e3c6dadSStefano Zampini if (nisdofs) { 56163e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 56173e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 56183e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 56193e3c6dadSStefano Zampini } 56203e3c6dadSStefano Zampini } 56213e3c6dadSStefano Zampini if (nisneu) { 56223e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 56233e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5624312be037SStefano Zampini } 562530368db7SStefano Zampini if (nisvert) { 562630368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 562730368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 562830368db7SStefano Zampini } 5629f9eb5b7dSStefano Zampini 5630f9eb5b7dSStefano Zampini /* get some info after set from options */ 5631f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5632f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 56334f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 56346e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5635f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5636f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5637f9eb5b7dSStefano Zampini } 563839f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 56394f3a063dSStefano Zampini if (isredundant) { 56404f3a063dSStefano Zampini KSP inner_ksp; 56414f3a063dSStefano Zampini PC inner_pc; 56424f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 56434f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 56444f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 56454f3a063dSStefano Zampini } 5646f9eb5b7dSStefano Zampini 5647f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5648fa7f1dd8SStefano Zampini if (coarse_reuse) { 564981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5650fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 56516e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5652fa7f1dd8SStefano Zampini } else { 56536e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5654fa7f1dd8SStefano Zampini } 5655c8587f34SStefano Zampini if (isbddc || isnn) { 565622bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 565770cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5658b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 565922b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 56606e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56616e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 56626e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 56636e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 566422b6e8a2SStefano Zampini } 566570cf5478SStefano Zampini } 566653a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 566770cf5478SStefano Zampini } else { 566822bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 566922bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 567022bc73bbSStefano Zampini } 567122bc73bbSStefano Zampini } else { 56722e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5673c8587f34SStefano Zampini } 5674c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5675c8587f34SStefano Zampini 56763301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 56775a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 56783301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 56793301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 56803301b35fSStefano Zampini } 56813301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 56823301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 56833301b35fSStefano Zampini } 56843301b35fSStefano Zampini if (pc->pmat->spd_set) { 56853301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 56863301b35fSStefano Zampini } 56876e683305SStefano Zampini /* set operators */ 56885f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 56896e683305SStefano Zampini if (pcbddc->dbg_flag) { 56906e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56916e683305SStefano Zampini } 56926e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 56936e683305SStefano Zampini coarse_mat = 0; 56946e683305SStefano Zampini } 56956e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5696b1ecc7b1SStefano Zampini #if 0 5697b9b85e73SStefano Zampini { 5698b9b85e73SStefano Zampini PetscViewer viewer; 5699b9b85e73SStefano Zampini char filename[256]; 5700b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5701b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5702b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5703b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5704b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5705b9b85e73SStefano Zampini } 5706b9b85e73SStefano Zampini #endif 5707c8587f34SStefano Zampini 5708c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5709298c0119SStefano Zampini #if 0 5710c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5711c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 571298a51de6SStefano Zampini } 5713298c0119SStefano Zampini #endif 5714b0f5fe93SStefano Zampini /* hack */ 571598a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 571698a51de6SStefano Zampini Vec crhs,csol; 571704708bb6SStefano Zampini 5718f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5719f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5720f347579bSStefano Zampini if (!csol) { 57212a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5722f9eb5b7dSStefano Zampini } 5723f347579bSStefano Zampini if (!crhs) { 57242a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5725f347579bSStefano Zampini } 5726b0f5fe93SStefano Zampini } 5727b0f5fe93SStefano Zampini 5728b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5729b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5730b0f5fe93SStefano Zampini 5731b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 57324f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 57334f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 57344f1b2e48SStefano Zampini } 5735b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5736b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5737b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5738b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5739b0f5fe93SStefano Zampini if (coarse_mat) { 5740b0f5fe93SStefano Zampini Vec nullv; 5741b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5742b0f5fe93SStefano Zampini PetscInt nl; 5743b0f5fe93SStefano Zampini 5744b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5745b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5746b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5747b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5748b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5749b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5750b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5751b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5752b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5753b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5754b0f5fe93SStefano Zampini } 5755b0f5fe93SStefano Zampini } 5756b0f5fe93SStefano Zampini 5757b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5758b0f5fe93SStefano Zampini PetscBool ispreonly; 5759b0f5fe93SStefano Zampini 5760b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5761b0f5fe93SStefano Zampini PetscBool isnull; 5762b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5763b0f5fe93SStefano Zampini if (isnull) { 576430368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5765b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5766b0f5fe93SStefano Zampini } else { 5767b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5768b0f5fe93SStefano Zampini } 5769b0f5fe93SStefano Zampini } else { 5770b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5771b0f5fe93SStefano Zampini } 5772b0f5fe93SStefano Zampini } 5773b0f5fe93SStefano Zampini /* setup coarse ksp */ 5774b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5775cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5776cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 57776e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5778c8587f34SStefano Zampini KSP check_ksp; 57792b510759SStefano Zampini KSPType check_ksp_type; 5780c8587f34SStefano Zampini PC check_pc; 57816e683305SStefano Zampini Vec check_vec,coarse_vec; 57826a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 57832b510759SStefano Zampini PetscInt its; 57846e683305SStefano Zampini PetscBool compute_eigs; 57856e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 57866e683305SStefano Zampini PetscInt neigs; 57878e185a42SStefano Zampini const char *prefix; 5788c8587f34SStefano Zampini 57892b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 57906e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5791422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 579223ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5793f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 57942b510759SStefano Zampini if (ispreonly) { 57952b510759SStefano Zampini check_ksp_type = KSPPREONLY; 57966e683305SStefano Zampini compute_eigs = PETSC_FALSE; 57972b510759SStefano Zampini } else { 5798cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 57996e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5800c8587f34SStefano Zampini } 5801c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 58026e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 58036e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 58046e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5805a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5806a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5807a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5808a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5809c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5810c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5811c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5812c8587f34SStefano Zampini /* create random vec */ 58136e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 58146e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5815c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5816c8587f34SStefano Zampini if (CoarseNullSpace) { 5817c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5818c8587f34SStefano Zampini } 58196e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5820c8587f34SStefano Zampini /* solve coarse problem */ 58216e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5822c8587f34SStefano Zampini if (CoarseNullSpace) { 58236e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5824c8587f34SStefano Zampini } 5825cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 58266e683305SStefano Zampini if (compute_eigs) { 5827854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5828854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 58296e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 58306e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 58316e683305SStefano Zampini lambda_min = eigs_r[0]; 58326e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 58336e683305SStefano Zampini if (lambda_max>lambda_min) { 5834cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5835cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5836cbcc2c2aSStefano Zampini } 5837c8587f34SStefano Zampini } 5838c8587f34SStefano Zampini } 5839cbcc2c2aSStefano Zampini 5840c8587f34SStefano Zampini /* check coarse problem residual error */ 58416e683305SStefano Zampini if (pcbddc->dbg_flag) { 58426e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 58436e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58446e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5845c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 58466e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 58476e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5848c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5849779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 58506e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 58516e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 58526e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 58536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5854b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5855b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5856b0f5fe93SStefano Zampini } 58576e683305SStefano Zampini if (compute_eigs) { 58586e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5859deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5860c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 58616e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 58626e683305SStefano 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); 58636e683305SStefano Zampini for (i=0;i<neigs;i++) { 58646e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5865c8587f34SStefano Zampini } 58666e683305SStefano Zampini } 58676e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 58686e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58696e683305SStefano Zampini } 5870c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 58716e683305SStefano Zampini if (compute_eigs) { 58726e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 58736e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5874c8587f34SStefano Zampini } 58756e683305SStefano Zampini } 58766e683305SStefano Zampini } 5877cbcc2c2aSStefano Zampini /* print additional info */ 5878cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 58796e683305SStefano Zampini /* waits until all processes reaches this point */ 58806e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5881cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5882cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5883cbcc2c2aSStefano Zampini } 5884cbcc2c2aSStefano Zampini 58852b510759SStefano Zampini /* free memory */ 5886c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5887fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5888c8587f34SStefano Zampini PetscFunctionReturn(0); 5889c8587f34SStefano Zampini } 5890674ae819SStefano Zampini 5891f34684f1SStefano Zampini #undef __FUNCT__ 5892f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5893f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5894f34684f1SStefano Zampini { 5895f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5896f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5897f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5898dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5899dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 590073be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5901dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5902f34684f1SStefano Zampini PetscErrorCode ierr; 5903f34684f1SStefano Zampini 5904f34684f1SStefano Zampini PetscFunctionBegin; 5905f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 59060e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 59070e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5908727cdba6SStefano Zampini } 5909dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 59103bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5911dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5912dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5913dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5914dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5915dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5916dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 59170e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 59180e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 59190e6343abSStefano Zampini } 5920dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5921dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5922dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5923dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5924dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5925f34684f1SStefano Zampini 5926f34684f1SStefano Zampini /* check numbering */ 5927f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5928019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5929dc456d91SStefano Zampini PetscInt i; 5930b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5931f34684f1SStefano Zampini 5932f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5933f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5934f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 59351575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5936019a44ceSStefano Zampini /* counter */ 5937019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5938019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5939019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5940019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5941019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5942019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5943f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5944f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5945727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5946f34684f1SStefano Zampini } 5947f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5948f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5949f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5950e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5951e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5952e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5953e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5954f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5955019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5956f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5957019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 595875c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 595975c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5960b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5961019a44ceSStefano 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); 5962f34684f1SStefano Zampini } 5963f34684f1SStefano Zampini } 5964019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5965b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5966f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5967f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5968f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5969f34684f1SStefano Zampini } 5970f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5971f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5972e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5973e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5974f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5975f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5976b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5977ca8b9ea9SStefano Zampini PetscInt *gidxs; 5978ca8b9ea9SStefano Zampini 5979ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 59803bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5981f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5982f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5983f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5984f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 59854bc2dc4bSStefano 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); 5986f34684f1SStefano Zampini } 5987f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5988ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5989f34684f1SStefano Zampini } 5990f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59911575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5992302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5993f34684f1SStefano Zampini } 59948bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5995f34684f1SStefano Zampini /* get back data */ 5996f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5997f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5998674ae819SStefano Zampini PetscFunctionReturn(0); 5999674ae819SStefano Zampini } 6000674ae819SStefano Zampini 6001e456f2a8SStefano Zampini #undef __FUNCT__ 6002e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6003a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6004e456f2a8SStefano Zampini { 6005e456f2a8SStefano Zampini IS localis_t; 6006a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6007e456f2a8SStefano Zampini PetscScalar *vals; 6008e456f2a8SStefano Zampini PetscErrorCode ierr; 6009e456f2a8SStefano Zampini 6010e456f2a8SStefano Zampini PetscFunctionBegin; 6011a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6012e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6013854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6014e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6015e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6016a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6017a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 60181035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6019a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 60201035eff8SStefano Zampini } 6021a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6022e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6023e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6024a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6025a7dc3881SStefano Zampini /* now compute set in local ordering */ 6026a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6027a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6028a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6029a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6030a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6031ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6032e456f2a8SStefano Zampini lsize++; 6033e456f2a8SStefano Zampini } 6034e456f2a8SStefano Zampini } 6035854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6036a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6037ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6038e456f2a8SStefano Zampini idxs[lsize++] = i; 6039e456f2a8SStefano Zampini } 6040e456f2a8SStefano Zampini } 6041a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6042a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6043e456f2a8SStefano Zampini *localis = localis_t; 6044e456f2a8SStefano Zampini PetscFunctionReturn(0); 6045e456f2a8SStefano Zampini } 6046906d46d4SStefano Zampini 6047906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6048906d46d4SStefano Zampini #undef __FUNCT__ 6049906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6050906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6051906d46d4SStefano Zampini { 6052906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6053906d46d4SStefano Zampini PetscErrorCode ierr; 6054906d46d4SStefano Zampini 6055906d46d4SStefano Zampini PetscFunctionBegin; 6056906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6057906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6058906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6059906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6060906d46d4SStefano Zampini PetscFunctionReturn(0); 6061906d46d4SStefano Zampini } 6062906d46d4SStefano Zampini 6063906d46d4SStefano Zampini #undef __FUNCT__ 6064906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6065906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6066906d46d4SStefano Zampini { 6067906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6068906d46d4SStefano Zampini PetscErrorCode ierr; 6069906d46d4SStefano Zampini 6070906d46d4SStefano Zampini PetscFunctionBegin; 6071906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6072906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6073906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6074906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6075906d46d4SStefano Zampini PetscFunctionReturn(0); 6076906d46d4SStefano Zampini } 6077b96c3477SStefano Zampini 6078b96c3477SStefano Zampini #undef __FUNCT__ 6079b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 608008122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6081b96c3477SStefano Zampini { 6082a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6083b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6084b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6085a64f4aa4SStefano Zampini Mat S_j; 6086b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6087b96c3477SStefano Zampini PetscBool free_used_adj; 6088b96c3477SStefano Zampini PetscErrorCode ierr; 6089b96c3477SStefano Zampini 6090b96c3477SStefano Zampini PetscFunctionBegin; 6091b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6092b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 609308122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6094b96c3477SStefano Zampini used_xadj = NULL; 6095b96c3477SStefano Zampini used_adjncy = NULL; 6096b96c3477SStefano Zampini } else { 609708122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 609808122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 609908122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 610008122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6101b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6102b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6103b96c3477SStefano Zampini } else { 61042fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6105b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6106b96c3477SStefano Zampini PetscInt nvtxs; 6107b96c3477SStefano Zampini 61082fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 61092fffb893SStefano Zampini if (flg_row) { 6110b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6111b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6112b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6113b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 61142fffb893SStefano Zampini } else { 61152fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 61162fffb893SStefano Zampini used_xadj = NULL; 61172fffb893SStefano Zampini used_adjncy = NULL; 61182fffb893SStefano Zampini } 61192fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6120b96c3477SStefano Zampini } 6121b96c3477SStefano Zampini } 6122d5574798SStefano Zampini 6123d5574798SStefano Zampini /* setup sub_schurs data */ 6124a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6125df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6126df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6127a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6128ca92afb2SStefano 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); 6129a64f4aa4SStefano Zampini } else { 61306816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 613104708bb6SStefano Zampini PetscBool isseqaij; 6132a3df083aSStefano Zampini PetscInt benign_n; 6133a3df083aSStefano Zampini 61345feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 61355feab87aSStefano Zampini PetscInt n_vertices; 61365feab87aSStefano Zampini 61375feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 61382034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 61395feab87aSStefano Zampini } 614004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 614104708bb6SStefano Zampini if (!isseqaij) { 614204708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 614304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 614404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 614504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 614604708bb6SStefano Zampini } else { 614704708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 614804708bb6SStefano Zampini } 614904708bb6SStefano Zampini } 6150a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6151a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6152ca92afb2SStefano Zampini } else { 6153a3df083aSStefano Zampini benign_n = 0; 6154ca92afb2SStefano Zampini } 6155a3df083aSStefano 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); 6156ca92afb2SStefano Zampini } 6157*d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6158b96c3477SStefano Zampini 6159b96c3477SStefano Zampini /* free adjacency */ 6160b96c3477SStefano Zampini if (free_used_adj) { 6161b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6162b96c3477SStefano Zampini } 6163b96c3477SStefano Zampini PetscFunctionReturn(0); 6164b96c3477SStefano Zampini } 6165b96c3477SStefano Zampini 6166b96c3477SStefano Zampini #undef __FUNCT__ 6167b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 616808122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6169b96c3477SStefano Zampini { 6170b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6171b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6172b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6173b96c3477SStefano Zampini PCBDDCGraph graph; 6174b96c3477SStefano Zampini PetscErrorCode ierr; 6175b96c3477SStefano Zampini 6176b96c3477SStefano Zampini PetscFunctionBegin; 6177b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 617808122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 61793301b35fSStefano Zampini IS verticesIS,verticescomm; 61803301b35fSStefano Zampini PetscInt vsize,*idxs; 6181b96c3477SStefano Zampini 6182b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 61833301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 61843301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61853301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 61863301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61873301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6188b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 61897fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 61903301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 61913301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6192b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6193b96c3477SStefano Zampini /* 6194b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6195b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6196b96c3477SStefano Zampini } 6197b96c3477SStefano Zampini */ 6198b96c3477SStefano Zampini } else { 6199b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6200b96c3477SStefano Zampini } 6201b96c3477SStefano Zampini 6202b96c3477SStefano Zampini /* sub_schurs init */ 6203a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6204a64f4aa4SStefano Zampini 6205b96c3477SStefano Zampini /* free graph struct */ 620608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6207b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6208b96c3477SStefano Zampini } 6209b96c3477SStefano Zampini PetscFunctionReturn(0); 6210b96c3477SStefano Zampini } 6211fa34dd3eSStefano Zampini 6212fa34dd3eSStefano Zampini #undef __FUNCT__ 6213fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6214fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6215fa34dd3eSStefano Zampini { 6216fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6217fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6218fa34dd3eSStefano Zampini PetscErrorCode ierr; 6219fa34dd3eSStefano Zampini 6220fa34dd3eSStefano Zampini PetscFunctionBegin; 6221fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6222fa34dd3eSStefano Zampini IS zerodiag = NULL; 62234f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6224fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 62254f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 622675c01103SStefano Zampini PetscReal norm; 6227fa34dd3eSStefano Zampini PetscInt i; 6228fa34dd3eSStefano Zampini 6229fa34dd3eSStefano Zampini /* B0 and B0_B */ 6230fa34dd3eSStefano Zampini if (zerodiag) { 6231fa34dd3eSStefano Zampini IS dummy; 6232fa34dd3eSStefano Zampini 62334f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 62344f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6235fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6236fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6237fa34dd3eSStefano Zampini } 6238fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6239fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6240fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6241fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6242fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6243fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6244fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6245fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6246fa34dd3eSStefano Zampini /* S_j */ 6247fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6248fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6249fa34dd3eSStefano Zampini 6250fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6251fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6252fa34dd3eSStefano Zampini /* continuous in primal space */ 6253fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6254fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6255fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6256fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 62574f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 62584f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6259fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6260fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6261fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6262fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6263fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6264fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6265fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6266fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6267fa34dd3eSStefano Zampini 6268fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6269fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6270fa34dd3eSStefano Zampini /* local with Schur */ 6271fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6272fa34dd3eSStefano Zampini if (zerodiag) { 6273fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62744f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6275fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6276fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6277fa34dd3eSStefano Zampini } 6278fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6279fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6280fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6281fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6282fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6283fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6284fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6285fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6286fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6287fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6288fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6289fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6290fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6291fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6292fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6293fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6294fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6295fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6296fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6297fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6298fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6299fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6300fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6301fa34dd3eSStefano Zampini if (zerodiag) { 6302fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6303fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 63044f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6305fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6306fa34dd3eSStefano Zampini } 6307fa34dd3eSStefano Zampini /* BDDC */ 6308fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6309fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6310fa34dd3eSStefano Zampini 6311fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6312fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6313fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6314fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 63154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 63164f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6317fa34dd3eSStefano Zampini } 63184f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6319fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6320fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6321fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6322fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6323fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6324fa34dd3eSStefano Zampini } 6325fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6326fa34dd3eSStefano Zampini } 6327