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; 135*df4d28bfSStefano 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; 4024f1b2e48SStefano Zampini PetscInt nz; 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); 4435408967cSStefano Zampini } 444dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4455408967cSStefano Zampini 4465408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4475408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4484f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4495408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4504f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4515408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4524f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4534f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4544f1b2e48SStefano 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); 4554f1b2e48SStefano Zampini } 4565408967cSStefano Zampini } 4575408967cSStefano Zampini PetscFunctionReturn(0); 4585408967cSStefano Zampini } 4595408967cSStefano Zampini 4605408967cSStefano Zampini #undef __FUNCT__ 461339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 462339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 463339f8db1SStefano Zampini { 464339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4654f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 466b0f5fe93SStefano Zampini PetscInt nz,n; 4674f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 468339f8db1SStefano Zampini PetscErrorCode ierr; 469339f8db1SStefano Zampini 470339f8db1SStefano Zampini PetscFunctionBegin; 4719f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 4729f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 473339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 474a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 475a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 476a3df083aSStefano Zampini } 477a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 478a3df083aSStefano Zampini pcbddc->benign_n = 0; 4794f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 4804f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 4814f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 4824f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 4834f1b2e48SStefano Zampini since the local Schur complements are SPD 4844f1b2e48SStefano Zampini */ 4854f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 4864f1b2e48SStefano Zampini have_null = PETSC_TRUE; 48740fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 4884f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 4894f1b2e48SStefano Zampini 4904f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 4914f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 4924f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 4934f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 494ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 49540fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 49640fa8d13SStefano Zampini if (!sorted) { 49740fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 49840fa8d13SStefano Zampini } 49940fa8d13SStefano Zampini } else { 50040fa8d13SStefano Zampini pressures = NULL; 50140fa8d13SStefano Zampini } 50297d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 50397d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 5044f1b2e48SStefano Zampini /* TODO: add check for shared dofs and raise error */ 50597d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 506339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 507339f8db1SStefano Zampini if (!sorted) { 508339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 509339f8db1SStefano Zampini } 510339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5114f1b2e48SStefano Zampini if (!nz) { 5124f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5134f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 51440fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 51540fa8d13SStefano Zampini } 5164f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5174f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5184f1b2e48SStefano Zampini zerodiag_subs = NULL; 5194f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5204f1b2e48SStefano Zampini if (has_null_pressures) { 5214f1b2e48SStefano Zampini IS *subs; 5224f1b2e48SStefano Zampini PetscInt nsubs,i; 5234f1b2e48SStefano Zampini 5244f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5254f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5264f1b2e48SStefano Zampini if (nsubs > 1) { 5274f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5284f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5294f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5304f1b2e48SStefano Zampini IS t_zerodiag_subs; 5314f1b2e48SStefano Zampini PetscInt nl; 5324f1b2e48SStefano Zampini 5334f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5344f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5354f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5364f1b2e48SStefano Zampini if (nl) { 5374f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5384f1b2e48SStefano Zampini 5394f1b2e48SStefano Zampini if (pressures) { 5404f1b2e48SStefano Zampini IS t_pressure_subs; 5414f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5424f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5434f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5444f1b2e48SStefano Zampini } 5454f1b2e48SStefano Zampini if (valid) { 5464f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5474f1b2e48SStefano Zampini pcbddc->benign_n++; 5484f1b2e48SStefano Zampini } else { 5494f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5504f1b2e48SStefano Zampini } 5514f1b2e48SStefano Zampini } 5524f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5534f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5544f1b2e48SStefano Zampini } 5554f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5564f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5574f1b2e48SStefano Zampini if (pressures) { 5584f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5594f1b2e48SStefano Zampini } 5604f1b2e48SStefano Zampini if (valid) { 5614f1b2e48SStefano Zampini pcbddc->benign_n = 1; 562ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5634f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5644f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 5654f1b2e48SStefano Zampini } 5664f1b2e48SStefano Zampini } 5674f1b2e48SStefano Zampini } 5684f1b2e48SStefano Zampini 5694f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 5704f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5714f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5724f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 5734f1b2e48SStefano Zampini have_null = PETSC_FALSE; 5744f1b2e48SStefano Zampini } 5754f1b2e48SStefano Zampini 5764f1b2e48SStefano Zampini /* final check for null pressures */ 5774f1b2e48SStefano Zampini if (zerodiag && pressures) { 5784f1b2e48SStefano Zampini PetscInt nz,np; 5794f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5804f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 5814f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 5824f1b2e48SStefano Zampini } 5834f1b2e48SStefano Zampini 5844f1b2e48SStefano Zampini if (recompute_zerodiag) { 5854f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 5864f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 5874f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 5894f1b2e48SStefano Zampini } else { 5904f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 5914f1b2e48SStefano Zampini 5924f1b2e48SStefano Zampini nzn = 0; 5934f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 5944f1b2e48SStefano Zampini PetscInt ns; 5954f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 5964f1b2e48SStefano Zampini nzn += ns; 5974f1b2e48SStefano Zampini } 5984f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 5994f1b2e48SStefano Zampini nzn = 0; 6004f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6014f1b2e48SStefano Zampini PetscInt ns,*idxs; 6024f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6034f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6044f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6054f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini nzn += ns; 6074f1b2e48SStefano Zampini } 6084f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6094f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6104f1b2e48SStefano Zampini } 6114f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6124f1b2e48SStefano Zampini } 6134f1b2e48SStefano Zampini 6144f1b2e48SStefano Zampini if (has_null_pressures) { 6154f1b2e48SStefano Zampini IS zerodiagc; 6164f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 6174f1b2e48SStefano Zampini PetscInt i,s,*nnz; 6184f1b2e48SStefano Zampini 6194f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 620339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 621339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 622339f8db1SStefano Zampini /* local change of basis for pressures */ 623339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 62497d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 625339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 626339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 627339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 6284f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 6294f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6304f1b2e48SStefano Zampini PetscInt nzs,j; 6314f1b2e48SStefano Zampini 6324f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 6334f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6344f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 6354f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 6364f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6374f1b2e48SStefano Zampini } 638339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 639339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 640339f8db1SStefano Zampini /* set identity on velocities */ 641339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 642339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 643339f8db1SStefano Zampini } 6444f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 6454f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 6469f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 6474f1b2e48SStefano 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); 648339f8db1SStefano Zampini /* set change on pressures */ 6494f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 6504f1b2e48SStefano Zampini PetscScalar *array; 6514f1b2e48SStefano Zampini PetscInt nzs; 6524f1b2e48SStefano Zampini 6534f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 6544f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 6554f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 656339f8db1SStefano Zampini PetscScalar vals[2]; 657339f8db1SStefano Zampini PetscInt cols[2]; 658339f8db1SStefano Zampini 659339f8db1SStefano Zampini cols[0] = idxs[i]; 6604f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 661339f8db1SStefano Zampini vals[0] = 1.; 662b0f5fe93SStefano Zampini vals[1] = 1.; 6634f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 664339f8db1SStefano Zampini } 6654f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 6664f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 6674f1b2e48SStefano Zampini array[nzs-1] = 1.; 6684f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 6694f1b2e48SStefano Zampini /* store local idxs for p0 */ 6704f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 6714f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 672339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 6734f1b2e48SStefano Zampini } 674339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 675339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 676a3df083aSStefano Zampini /* project if needed */ 677a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 67897d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 679339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 68097d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 68197d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 682a3df083aSStefano Zampini } 6834f1b2e48SStefano Zampini /* store global idxs for p0 */ 6844f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 685339f8db1SStefano Zampini } 686ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 6874f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 688b0f5fe93SStefano Zampini 689b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 690b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 691339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 692339f8db1SStefano Zampini PetscFunctionReturn(0); 693339f8db1SStefano Zampini } 694339f8db1SStefano Zampini 695339f8db1SStefano Zampini #undef __FUNCT__ 696015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 697015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 698efc2fbd9SStefano Zampini { 699efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 700efc2fbd9SStefano Zampini PetscErrorCode ierr; 701efc2fbd9SStefano Zampini 702efc2fbd9SStefano Zampini PetscFunctionBegin; 703efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 704efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 7054f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 706efc2fbd9SStefano Zampini } 707015636ebSStefano Zampini if (get) { /* use SF to get values */ 708efc2fbd9SStefano Zampini PetscScalar *array; 709efc2fbd9SStefano Zampini 710efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7114f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 7124f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 713efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7144f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 7154f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 716efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 717efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 718efc2fbd9SStefano Zampini } 719efc2fbd9SStefano Zampini PetscFunctionReturn(0); 720efc2fbd9SStefano Zampini } 721efc2fbd9SStefano Zampini 722efc2fbd9SStefano Zampini #undef __FUNCT__ 723c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 724c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 725c263805aSStefano Zampini { 726c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 727c263805aSStefano Zampini PetscErrorCode ierr; 728c263805aSStefano Zampini 729c263805aSStefano Zampini PetscFunctionBegin; 730c263805aSStefano Zampini /* TODO: add error checking 731c263805aSStefano Zampini - avoid nested pop (or push) calls. 732c263805aSStefano Zampini - cannot push before pop. 7331c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 734c263805aSStefano Zampini */ 7354f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 736efc2fbd9SStefano Zampini PetscFunctionReturn(0); 737efc2fbd9SStefano Zampini } 738c263805aSStefano Zampini if (pop) { 739a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7404f1b2e48SStefano Zampini IS is_p0; 7414f1b2e48SStefano Zampini MatReuse reuse; 742c263805aSStefano Zampini 743c263805aSStefano Zampini /* extract B_0 */ 7444f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 7454f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 7464f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 7474f1b2e48SStefano Zampini } 7484f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 7494f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 750c263805aSStefano Zampini /* remove rows and cols from local problem */ 751c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 75297d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 7534f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 7544f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 755a3df083aSStefano Zampini } else { 756a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 757a3df083aSStefano Zampini PetscScalar *vals; 758a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 759a3df083aSStefano Zampini 760a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 761a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 762a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 763a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 764a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 765a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 766a3df083aSStefano Zampini /* this matrix is very sparse: the nnz pattern is not known unless we do 2 sweeps of the next loop. 767a3df083aSStefano Zampini Setting nnz=100 should be more than enough */ 768a3df083aSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,100,NULL);CHKERRQ(ierr); 769a3df083aSStefano Zampini } 770a3df083aSStefano Zampini 771a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 772a3df083aSStefano Zampini PetscScalar *array; 773a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 774a3df083aSStefano Zampini 775a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 776a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 777a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 778a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 779a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 780a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 781a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 782a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 783a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 784a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 785a3df083aSStefano Zampini cum = 0; 786a3df083aSStefano Zampini for (j=0;j<n;j++) { 787a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 788a3df083aSStefano Zampini vals[cum] = array[j]; 789a3df083aSStefano Zampini idxs_ins[cum] = j; 790a3df083aSStefano Zampini cum++; 791a3df083aSStefano Zampini } 792a3df083aSStefano Zampini } 793a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 794a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 795a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 796a3df083aSStefano Zampini } 797a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 798a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 799a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 800a3df083aSStefano Zampini } 801c263805aSStefano Zampini } else { /* push */ 802a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8034f1b2e48SStefano Zampini PetscInt i; 8044f1b2e48SStefano Zampini 8054f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8064f1b2e48SStefano Zampini PetscScalar *B0_vals; 8074f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 8084f1b2e48SStefano Zampini 8094f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8104f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8117b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8124f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 8134f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8144f1b2e48SStefano Zampini } 815c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 816c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 817a3df083aSStefano Zampini } else { 818a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 819a3df083aSStefano Zampini } 820c263805aSStefano Zampini } 821c263805aSStefano Zampini PetscFunctionReturn(0); 822c263805aSStefano Zampini } 823c263805aSStefano Zampini 824c263805aSStefano Zampini #undef __FUNCT__ 825b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 82608122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 827b1b3d7a2SStefano Zampini { 828b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 82908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 83008122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 83108122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 83208122e43SStefano Zampini PetscScalar *work,lwork; 83308122e43SStefano Zampini PetscScalar *St,*S,*eigv; 83408122e43SStefano Zampini PetscScalar *Sarray,*Starray; 83508122e43SStefano Zampini PetscReal *eigs,thresh; 8361b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 837f6f667cfSStefano Zampini PetscBool allocated_S_St; 83808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 83908122e43SStefano Zampini PetscReal *rwork; 84008122e43SStefano Zampini #endif 841b1b3d7a2SStefano Zampini PetscErrorCode ierr; 842b1b3d7a2SStefano Zampini 843b1b3d7a2SStefano Zampini PetscFunctionBegin; 844*df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 845*df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 84608122e43SStefano Zampini } 84708122e43SStefano Zampini 84806a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 84906a4e24aSStefano 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); 85006a4e24aSStefano Zampini } 85106a4e24aSStefano Zampini 852fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 853fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 854fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 855fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 8561575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 857fd14bc51SStefano Zampini } 858fd14bc51SStefano Zampini 859e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 860e496cd5dSStefano 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); 861e496cd5dSStefano Zampini } 862e496cd5dSStefano Zampini 86308122e43SStefano Zampini /* max size of subsets */ 86408122e43SStefano Zampini mss = 0; 86508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 86608122e43SStefano Zampini PetscInt subset_size; 867862806e4SStefano Zampini 86808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 86908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 87008122e43SStefano Zampini } 87108122e43SStefano Zampini 87208122e43SStefano Zampini /* min/max and threshold */ 87308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 874f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 87508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 876f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 877f6f667cfSStefano Zampini if (nmin) { 878f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 879f6f667cfSStefano Zampini } 88008122e43SStefano Zampini 88108122e43SStefano Zampini /* allocate lapack workspace */ 88208122e43SStefano Zampini cum = cum2 = 0; 88308122e43SStefano Zampini maxneigs = 0; 88408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 88508122e43SStefano Zampini PetscInt n,subset_size; 886f6f667cfSStefano Zampini 88708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 88808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 8899162d606SStefano Zampini cum += subset_size; 8909162d606SStefano Zampini cum2 += subset_size*n; 89108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 89208122e43SStefano Zampini } 89308122e43SStefano Zampini if (mss) { 8949ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 89508122e43SStefano Zampini PetscBLASInt B_itype = 1; 89608122e43SStefano Zampini PetscBLASInt B_N = mss; 8974c6709b3SStefano Zampini PetscReal zero = 0.0; 8984c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 89908122e43SStefano Zampini 90008122e43SStefano Zampini B_lwork = -1; 90108122e43SStefano Zampini S = NULL; 90208122e43SStefano Zampini St = NULL; 903a58a30b4SStefano Zampini eigs = NULL; 904a58a30b4SStefano Zampini eigv = NULL; 905a58a30b4SStefano Zampini B_iwork = NULL; 906a58a30b4SStefano Zampini B_ifail = NULL; 907d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 908d1710679SStefano Zampini rwork = NULL; 909d1710679SStefano Zampini #endif 9108bec7fa6SStefano Zampini thresh = 1.0; 91108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 91208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 91308122e43SStefano 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)); 91408122e43SStefano Zampini #else 91508122e43SStefano 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)); 91608122e43SStefano Zampini #endif 91708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 91808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 91908122e43SStefano Zampini } else { 92008122e43SStefano Zampini /* TODO */ 92108122e43SStefano Zampini } 92208122e43SStefano Zampini } else { 92308122e43SStefano Zampini lwork = 0; 92408122e43SStefano Zampini } 92508122e43SStefano Zampini 92608122e43SStefano Zampini nv = 0; 927d62866d3SStefano 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) */ 928d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 92908122e43SStefano Zampini } 9304c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 931f6f667cfSStefano Zampini if (allocated_S_St) { 932f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 933f6f667cfSStefano Zampini } 934f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 93508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 93608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 93708122e43SStefano Zampini #endif 9389162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 9399162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 9409162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 94108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 9429162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 94308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 94408122e43SStefano Zampini 94508122e43SStefano Zampini maxneigs = 0; 94608122e43SStefano Zampini cum = cum2 = cumarray = 0; 9479162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 9489162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 949d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 95008122e43SStefano Zampini const PetscInt *idxs; 95108122e43SStefano Zampini 952d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 95308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 95408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 95508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 95608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 9579162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 9589162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 95908122e43SStefano Zampini } 96008122e43SStefano Zampini cum2 = cum; 961d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 96208122e43SStefano Zampini } 96308122e43SStefano Zampini 96408122e43SStefano Zampini if (mss) { /* multilevel */ 96508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 96608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 96708122e43SStefano Zampini } 96808122e43SStefano Zampini 969ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 97008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 97108122e43SStefano Zampini const PetscInt *idxs; 972f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 973862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 97408122e43SStefano Zampini PetscBLASInt B_N; 975aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 97608122e43SStefano Zampini 977862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 978ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 979f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 980f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 9819ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 982aff50787SStefano Zampini PetscInt j,k; 983aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 984aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 985aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 98608122e43SStefano Zampini } 98708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 988aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 989aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 990aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 991aff50787SStefano Zampini } 99208122e43SStefano Zampini } 99308122e43SStefano Zampini } else { 99408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 99508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 99608122e43SStefano Zampini } 9978bec7fa6SStefano Zampini } else { 998f6f667cfSStefano Zampini S = Sarray + cumarray; 999f6f667cfSStefano Zampini St = Starray + cumarray; 10008bec7fa6SStefano Zampini } 100108122e43SStefano Zampini 1002aff50787SStefano Zampini /* see if we can save some work */ 1003aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1004aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1005aff50787SStefano Zampini } 1006aff50787SStefano Zampini 1007aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1008aff50787SStefano Zampini B_neigs = 0; 1009aff50787SStefano Zampini } else { 10109ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 101108122e43SStefano Zampini PetscBLASInt B_itype = 1; 1012f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 10134c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 10149552c7c7SStefano Zampini PetscInt nmin_s; 101508122e43SStefano Zampini 1016fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 10178bec7fa6SStefano 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]]); 1018fd14bc51SStefano Zampini } 1019d16cbb6bSStefano Zampini 102008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1021d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1022d16cbb6bSStefano Zampini 1023d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 102408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1025f6f667cfSStefano 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)); 102608122e43SStefano Zampini #else 1027f6f667cfSStefano 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)); 102808122e43SStefano Zampini #endif 1029d16cbb6bSStefano Zampini } else { 1030d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1031d16cbb6bSStefano Zampini B_IL = 1; 1032d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1033d16cbb6bSStefano 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)); 1034d16cbb6bSStefano Zampini #else 1035d16cbb6bSStefano 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)); 1036d16cbb6bSStefano Zampini #endif 1037d16cbb6bSStefano Zampini } 103808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 103908122e43SStefano Zampini if (B_ierr) { 104008122e43SStefano Zampini if (B_ierr < 0 ) { 104108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 104208122e43SStefano Zampini } else if (B_ierr <= B_N) { 104308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 104408122e43SStefano Zampini } else { 10459552c7c7SStefano 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); 104608122e43SStefano Zampini } 104708122e43SStefano Zampini } 104808122e43SStefano Zampini 104908122e43SStefano Zampini if (B_neigs > nmax) { 1050fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1051fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1052fd14bc51SStefano Zampini } 1053f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 105408122e43SStefano Zampini B_neigs = nmax; 105508122e43SStefano Zampini } 105608122e43SStefano Zampini 10579552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 10589552c7c7SStefano Zampini if (B_neigs < nmin_s) { 105908122e43SStefano Zampini PetscBLASInt B_neigs2; 106008122e43SStefano Zampini 1061f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1062f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1063fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1064fd14bc51SStefano 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); 1065fd14bc51SStefano Zampini } 10669ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 106708122e43SStefano Zampini PetscInt j; 106808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 106908122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 107008122e43SStefano Zampini } 107108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 107208122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 107308122e43SStefano Zampini } 107408122e43SStefano Zampini } else { 107508122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 107608122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 107708122e43SStefano Zampini } 107808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 107908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1080f6f667cfSStefano 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)); 108108122e43SStefano Zampini #else 1082f6f667cfSStefano 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)); 108308122e43SStefano Zampini #endif 108408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 108508122e43SStefano Zampini B_neigs += B_neigs2; 108608122e43SStefano Zampini } 108708122e43SStefano Zampini if (B_ierr) { 108808122e43SStefano Zampini if (B_ierr < 0 ) { 108908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 109008122e43SStefano Zampini } else if (B_ierr <= B_N) { 109108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 109208122e43SStefano Zampini } else { 10939552c7c7SStefano 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); 109408122e43SStefano Zampini } 109508122e43SStefano Zampini } 1096fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1097ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 109808122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 109908122e43SStefano Zampini if (eigs[j] == 0.0) { 1100ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 110108122e43SStefano Zampini } else { 1102ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1103fd14bc51SStefano Zampini } 110408122e43SStefano Zampini } 110508122e43SStefano Zampini } 110608122e43SStefano Zampini } else { 110708122e43SStefano Zampini /* TODO */ 110808122e43SStefano Zampini } 1109aff50787SStefano Zampini } 11108bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 11118bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 11129162d606SStefano Zampini if (B_neigs) { 11139162d606SStefano 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); 1114fd14bc51SStefano Zampini 1115fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 11169552c7c7SStefano Zampini PetscInt ii; 11179552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1118ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11199552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1120ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1121ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1122ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1123ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1124ac47001eSStefano Zampini #else 1125ac47001eSStefano 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); 1126ac47001eSStefano Zampini #endif 11279552c7c7SStefano Zampini } 11289552c7c7SStefano Zampini } 1129fd14bc51SStefano Zampini } 113008122e43SStefano Zampini #if 0 11319162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 113208122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 113308122e43SStefano Zampini PetscScalar norm; 113408122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 11359162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 11369162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 113708122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 113808122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 113908122e43SStefano Zampini } else { 114008122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 114108122e43SStefano Zampini } 11429162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 1143b1b3d7a2SStefano Zampini } 1144b1b3d7a2SStefano Zampini #endif 11459162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 11469162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 11479162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 11489162d606SStefano Zampini cum++; 114908122e43SStefano Zampini } 115008122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 115108122e43SStefano Zampini /* shift for next computation */ 115208122e43SStefano Zampini cumarray += subset_size*subset_size; 115308122e43SStefano Zampini } 1154fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1155fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1156fd14bc51SStefano Zampini } 115708122e43SStefano Zampini 115808122e43SStefano Zampini if (mss) { 115908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 116008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1161f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1162f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1163f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 116408122e43SStefano Zampini } 1165f6f667cfSStefano Zampini if (allocated_S_St) { 1166f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1167f6f667cfSStefano Zampini } 1168f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 116908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 117008122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 117108122e43SStefano Zampini #endif 117208122e43SStefano Zampini if (pcbddc->dbg_flag) { 11731b968477SStefano Zampini PetscInt maxneigs_r; 117408122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 11759b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 117608122e43SStefano Zampini } 117708122e43SStefano Zampini PetscFunctionReturn(0); 117808122e43SStefano Zampini } 1179b1b3d7a2SStefano Zampini 1180674ae819SStefano Zampini #undef __FUNCT__ 1181c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1182c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1183c8587f34SStefano Zampini { 1184c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 11858629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1186c8587f34SStefano Zampini PetscErrorCode ierr; 1187c8587f34SStefano Zampini 1188c8587f34SStefano Zampini PetscFunctionBegin; 1189f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 11905e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1191c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1192c8587f34SStefano Zampini 1193684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 11940fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1195684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1196c8587f34SStefano Zampini 1197c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1198b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1199c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1200c8587f34SStefano Zampini } 1201c8587f34SStefano Zampini 12028629588bSStefano Zampini /* 12038629588bSStefano Zampini Setup local correction and local part of coarse basis. 12048629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12058629588bSStefano Zampini */ 120647f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12078629588bSStefano Zampini 12088629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12098629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12108629588bSStefano Zampini 12118629588bSStefano Zampini /* free */ 12128629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1213c8587f34SStefano Zampini PetscFunctionReturn(0); 1214c8587f34SStefano Zampini } 1215c8587f34SStefano Zampini 1216c8587f34SStefano Zampini #undef __FUNCT__ 1217674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1218674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1219674ae819SStefano Zampini { 1220674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1221674ae819SStefano Zampini PetscErrorCode ierr; 1222674ae819SStefano Zampini 1223674ae819SStefano Zampini PetscFunctionBegin; 1224674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1225674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 122630368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1227674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1228674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1229785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1230674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1231f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1232f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1233785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 123463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 123563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1236674ae819SStefano Zampini PetscFunctionReturn(0); 1237674ae819SStefano Zampini } 1238674ae819SStefano Zampini 1239674ae819SStefano Zampini #undef __FUNCT__ 1240674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1241674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1242674ae819SStefano Zampini { 1243674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12444f1b2e48SStefano Zampini PetscInt i; 1245674ae819SStefano Zampini PetscErrorCode ierr; 1246674ae819SStefano Zampini 1247674ae819SStefano Zampini PetscFunctionBegin; 1248b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1249674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1250674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1251674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 12524f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 12534f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 12544f1b2e48SStefano Zampini } 12554f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1256b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1257674ae819SStefano Zampini PetscFunctionReturn(0); 1258674ae819SStefano Zampini } 1259674ae819SStefano Zampini 1260674ae819SStefano Zampini #undef __FUNCT__ 1261674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1262674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1263674ae819SStefano Zampini { 1264674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1265674ae819SStefano Zampini PetscErrorCode ierr; 1266674ae819SStefano Zampini 1267674ae819SStefano Zampini PetscFunctionBegin; 1268674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 126958da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1270ca92afb2SStefano Zampini PetscScalar *array; 127106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 127206656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 127358da7f69SStefano Zampini } 1274674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1275674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 127615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 127715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1278674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1279674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1280674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 128106656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1282674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1283674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 12848ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1285674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1286674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1287674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1288f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1289f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1290f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1291f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1292727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 12930e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1294f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 129570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 12966e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 129781d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 12980369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 12998b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 13004f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13018b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1302ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1303ca92afb2SStefano Zampini PetscInt i; 1304ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1305ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1306ca92afb2SStefano Zampini } 1307ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1308ca92afb2SStefano Zampini } 13094f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1310674ae819SStefano Zampini PetscFunctionReturn(0); 1311674ae819SStefano Zampini } 1312674ae819SStefano Zampini 1313674ae819SStefano Zampini #undef __FUNCT__ 1314f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1315f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 13166bfb1811SStefano Zampini { 13176bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13186bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 13196bfb1811SStefano Zampini VecType impVecType; 13204f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 13216bfb1811SStefano Zampini PetscErrorCode ierr; 13226bfb1811SStefano Zampini 13236bfb1811SStefano Zampini PetscFunctionBegin; 1324f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1325019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1326f4ddd8eeSStefano Zampini } 1327e7b262bdSStefano Zampini /* get sizes */ 13284f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1329b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 13306bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1331e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1332e7b262bdSStefano Zampini /* R nodes */ 1333e7b262bdSStefano Zampini old_size = -1; 1334e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1335e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1336e7b262bdSStefano Zampini } 1337e7b262bdSStefano Zampini if (n_R != old_size) { 1338e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1339e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13406bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 13416bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 13426bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 13436bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1344e7b262bdSStefano Zampini } 1345e7b262bdSStefano Zampini /* local primal dofs */ 1346e7b262bdSStefano Zampini old_size = -1; 1347e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1348e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1349e7b262bdSStefano Zampini } 1350e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1351e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 135283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1353e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 13546bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1355e7b262bdSStefano Zampini } 1356e7b262bdSStefano Zampini /* local explicit constraints */ 1357e7b262bdSStefano Zampini old_size = -1; 1358e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1359e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1360e7b262bdSStefano Zampini } 1361e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1362e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 136383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 136483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 136583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 136683b7ccabSStefano Zampini } 13676bfb1811SStefano Zampini PetscFunctionReturn(0); 13686bfb1811SStefano Zampini } 13696bfb1811SStefano Zampini 13706bfb1811SStefano Zampini #undef __FUNCT__ 137147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 137247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 137388ebb749SStefano Zampini { 137425084f0cSStefano Zampini PetscErrorCode ierr; 137525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 137688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 137788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1378d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 137925084f0cSStefano Zampini /* submatrices of local problem */ 138080677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1381019a44ceSStefano Zampini /* submatrices of benign trick */ 1382d16cbb6bSStefano Zampini Mat B0_V = NULL; 138306656605SStefano Zampini /* submatrices of local coarse problem */ 138406656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 138525084f0cSStefano Zampini /* working matrices */ 138606656605SStefano Zampini Mat C_CR; 138725084f0cSStefano Zampini /* additional working stuff */ 138806656605SStefano Zampini PC pc_R; 13894f1b2e48SStefano Zampini Mat F; 1390a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 139106656605SStefano Zampini 139225084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 139306656605SStefano Zampini PetscScalar *work; 139406656605SStefano Zampini PetscInt *idx_V_B; 1395ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 139606656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1397ffd830a3SStefano Zampini 139825084f0cSStefano Zampini /* some shortcuts to scalars */ 139906656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 140088ebb749SStefano Zampini 140188ebb749SStefano Zampini PetscFunctionBegin; 1402ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1403ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1404ffd830a3SStefano Zampini } 1405ffd830a3SStefano Zampini 1406ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1407b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14084f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1409b371cd4fSStefano Zampini n_B = pcis->n_B; 1410b371cd4fSStefano Zampini n_D = pcis->n - n_B; 141188ebb749SStefano Zampini n_R = pcis->n - n_vertices; 141288ebb749SStefano Zampini 141388ebb749SStefano Zampini /* vertices in boundary numbering */ 1414785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 14150e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 141688ebb749SStefano Zampini if (i != n_vertices) { 141722d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 141888ebb749SStefano Zampini } 141988ebb749SStefano Zampini 142006656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1421019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 142206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 142306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 142406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 142506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 142606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 142706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 142806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 142906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 143006656605SStefano Zampini 143106656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 143206656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 143306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 143406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 143506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1436ffd830a3SStefano Zampini lda_rhs = n_R; 1437a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 143806656605SStefano Zampini if (isLU || isILU || isCHOL) { 143906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1440*df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1441*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1442d62866d3SStefano Zampini MatFactorType type; 1443d62866d3SStefano Zampini 1444*df4d28bfSStefano Zampini F = reuse_solver->F; 14456816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1446d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1447ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1448*df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 144906656605SStefano Zampini } else { 145006656605SStefano Zampini F = NULL; 145106656605SStefano Zampini } 145206656605SStefano Zampini 1453ffd830a3SStefano Zampini /* allocate workspace */ 1454ffd830a3SStefano Zampini n = 0; 1455ffd830a3SStefano Zampini if (n_constraints) { 1456ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1457ffd830a3SStefano Zampini } 1458ffd830a3SStefano Zampini if (n_vertices) { 1459ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1460ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1461ffd830a3SStefano Zampini } 1462ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1463ffd830a3SStefano Zampini 146488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 146588ebb749SStefano Zampini if (n_constraints) { 146606656605SStefano Zampini Mat M1,M2,M3; 146780677318SStefano Zampini Mat auxmat; 146806656605SStefano Zampini IS is_aux; 146980677318SStefano Zampini PetscScalar *array,*array2; 147006656605SStefano Zampini 1471f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 147280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 147388ebb749SStefano Zampini 147425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 147525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 14768ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 147780677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 147888ebb749SStefano Zampini 147980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 148080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1481ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 148288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 148306656605SStefano Zampini const PetscScalar *row_cmat_values; 148406656605SStefano Zampini const PetscInt *row_cmat_indices; 148506656605SStefano Zampini PetscInt size_of_constraint,j; 148688ebb749SStefano Zampini 148706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 148806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1489ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 149006656605SStefano Zampini } 149106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 149206656605SStefano Zampini } 1493ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 149406656605SStefano Zampini if (F) { 149506656605SStefano Zampini Mat B; 149606656605SStefano Zampini 1497ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1498a3df083aSStefano Zampini if (need_benign_correction) { 1499a3df083aSStefano Zampini PetscScalar *marr; 1500*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1501a3df083aSStefano Zampini 1502a3df083aSStefano Zampini ierr = MatDenseGetArray(B,&marr);CHKERRQ(ierr); 1503a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1504a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1505*df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1506a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1507a3df083aSStefano Zampini } 1508a3df083aSStefano Zampini ierr = MatDenseRestoreArray(B,&marr);CHKERRQ(ierr); 1509a3df083aSStefano Zampini } 151080677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1511a3df083aSStefano Zampini if (need_benign_correction) { 1512a3df083aSStefano Zampini PetscScalar *marr; 1513*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1514a3df083aSStefano Zampini 1515a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1516a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1517a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1518*df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1519a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1520a3df083aSStefano Zampini } 1521a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1522a3df083aSStefano Zampini } 152306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 152406656605SStefano Zampini } else { 152580677318SStefano Zampini PetscScalar *marr; 152680677318SStefano Zampini 152780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 152806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1529ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1530ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 153106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 153206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 153306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 153406656605SStefano Zampini } 153580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 153606656605SStefano Zampini } 153780677318SStefano Zampini if (!pcbddc->switch_static) { 153880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 153980677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 154080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 154180677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1542ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 154380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 154480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 154580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 154680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 154780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 154880677318SStefano Zampini } 154980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 155080677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 155180677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 155280677318SStefano Zampini } else { 1553ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1554ffd830a3SStefano Zampini IS dummy; 1555ffd830a3SStefano Zampini 1556ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1557ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1558ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1559ffd830a3SStefano Zampini } else { 156080677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 156180677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1562ffd830a3SStefano Zampini } 156325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 156480677318SStefano Zampini } 156580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 156680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 156780677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 156806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 156906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 157080677318SStefano Zampini if (isCHOL) { 157180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 157280677318SStefano Zampini } else { 157325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 157480677318SStefano Zampini } 157580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 157606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 157725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 157825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 157925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 158080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 158180677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 158280677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 158306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 158406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1585f4ddd8eeSStefano Zampini } 158688ebb749SStefano Zampini /* Get submatrices from subdomain matrix */ 15874f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1588d16cbb6bSStefano Zampini IS dummy; 1589d16cbb6bSStefano Zampini Mat B0_R; 1590d16cbb6bSStefano Zampini PetscReal norm; 1591d16cbb6bSStefano Zampini 15924f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 15934f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1594d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1595d16cbb6bSStefano Zampini if (norm > PETSC_SMALL) { 1596d16cbb6bSStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! ||B0_R|| = %f (should be numerically 0.)",norm); 1597d16cbb6bSStefano Zampini } 1598d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1599d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1600d16cbb6bSStefano Zampini } 1601d16cbb6bSStefano Zampini 160288ebb749SStefano Zampini if (n_vertices) { 160306656605SStefano Zampini IS is_aux; 16043a50541eSStefano Zampini 1605*df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16066816873aSStefano Zampini IS tis; 16076816873aSStefano Zampini 16086816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16096816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16106816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16116816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16126816873aSStefano Zampini } else { 16133a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16146816873aSStefano Zampini } 16159577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16169577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 161704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 16184f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1619019a44ceSStefano Zampini IS dummy; 1620019a44ceSStefano Zampini 16214f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16224f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1623019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1624019a44ceSStefano Zampini } 162525084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 162688ebb749SStefano Zampini } 162788ebb749SStefano Zampini 162888ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1629f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 163006656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 163106656605SStefano Zampini if (pcbddc->coarse_phi_D) { 163206656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 163306656605SStefano Zampini } 1634f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 163506656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 163606656605SStefano Zampini PetscScalar *marray; 163706656605SStefano Zampini 163806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 163906656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1640f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1641f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1642f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1643f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1644f4ddd8eeSStefano Zampini } 1645f4ddd8eeSStefano Zampini } 164606656605SStefano Zampini 1647f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 164806656605SStefano Zampini PetscScalar *marray; 164988ebb749SStefano Zampini 165006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16518eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 165206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 165388ebb749SStefano Zampini } 16543301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 165506656605SStefano Zampini n *= 2; 165688ebb749SStefano Zampini } 165706656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 165806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 165906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16608eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 166106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 166206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 166388ebb749SStefano Zampini } 16643301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 166506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 16668eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 166706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 166806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 166988ebb749SStefano Zampini } 167088ebb749SStefano Zampini } else { 1671c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1672c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 16731b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1674c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1675c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1676c0553b1fSStefano Zampini } 167788ebb749SStefano Zampini } 167806656605SStefano Zampini } 1679019a44ceSStefano Zampini 168006656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 16814f1b2e48SStefano Zampini p0_lidx_I = NULL; 16824f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1683d12edf2fSStefano Zampini const PetscInt *idxs; 1684d12edf2fSStefano Zampini 1685d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 16864f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 16874f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 16884f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 16894f1b2e48SStefano Zampini } 1690d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1691d12edf2fSStefano Zampini } 1692d16cbb6bSStefano Zampini 169306656605SStefano Zampini /* vertices */ 169406656605SStefano Zampini if (n_vertices) { 169516f15bc4SStefano Zampini 1696ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 169704708bb6SStefano Zampini 169816f15bc4SStefano Zampini if (n_R) { 169914393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 170006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 170116f15bc4SStefano Zampini PetscScalar *x,*y; 170204708bb6SStefano Zampini PetscBool isseqaij; 170306656605SStefano Zampini 170421eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 170514393ed6SStefano Zampini if (need_benign_correction) { 170614393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 170714393ed6SStefano Zampini IS is_p0; 170814393ed6SStefano Zampini PetscInt *idxs_p0,n; 170914393ed6SStefano Zampini 171014393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 171114393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 171214393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 171314393ed6SStefano Zampini if (n != pcbddc->benign_n) { 171414393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 171514393ed6SStefano Zampini } 171614393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 171714393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 171814393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 171914393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 172014393ed6SStefano Zampini } 172114393ed6SStefano Zampini 1722ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1723ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1724ffd830a3SStefano Zampini } else { 1725ca92afb2SStefano Zampini PetscScalar *av,*array; 1726ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1727ca92afb2SStefano Zampini PetscInt n; 1728ca92afb2SStefano Zampini PetscBool flg_row; 1729ffd830a3SStefano Zampini 1730ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1731ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1732ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1733ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1734ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1735ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1736ca92afb2SStefano Zampini PetscInt j; 1737ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1738ffd830a3SStefano Zampini } 1739ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1740ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1741ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1742ffd830a3SStefano Zampini } 1743ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 174414393ed6SStefano Zampini if (F) { 1745a3df083aSStefano Zampini if (need_benign_correction) { 1746*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1747a3df083aSStefano Zampini PetscScalar *marr; 1748a3df083aSStefano Zampini 1749a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 175014393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 175114393ed6SStefano Zampini 175214393ed6SStefano Zampini | 0 0 0 | (V) 175314393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 175414393ed6SStefano Zampini | 0 0 -1 | (p0) 175514393ed6SStefano Zampini 175614393ed6SStefano Zampini */ 1757*df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 175814393ed6SStefano Zampini const PetscScalar *vals; 175914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 176014393ed6SStefano Zampini PetscInt n,j,nz; 176114393ed6SStefano Zampini 1762*df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1763*df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 176414393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 176514393ed6SStefano Zampini for (j=0;j<n;j++) { 176614393ed6SStefano Zampini PetscScalar val = vals[j]; 176714393ed6SStefano Zampini PetscInt k,col = idxs[j]; 176814393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 176914393ed6SStefano Zampini } 177014393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1771*df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 177214393ed6SStefano Zampini } 177314393ed6SStefano Zampini /* need to correct the rhs */ 1774a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1775a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1776*df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1777a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1778a3df083aSStefano Zampini } 1779a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1780a3df083aSStefano Zampini } 178106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 178214393ed6SStefano Zampini /* need to correct the solution */ 1783a3df083aSStefano Zampini if (need_benign_correction) { 1784*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1785a3df083aSStefano Zampini PetscScalar *marr; 1786a3df083aSStefano Zampini 1787a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1788a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1789a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1790*df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1791a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1792a3df083aSStefano Zampini } 1793a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1794a3df083aSStefano Zampini } 179506656605SStefano Zampini } else { 179606656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 179706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1798ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1799ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 180006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 180106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 180206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 180306656605SStefano Zampini } 180406656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 180506656605SStefano Zampini } 180680677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1807ffd830a3SStefano Zampini /* S_VV and S_CV */ 180806656605SStefano Zampini if (n_constraints) { 180906656605SStefano Zampini Mat B; 181080677318SStefano Zampini 1811ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 181280677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1813ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1814ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 181580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 181780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 181880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 181980677318SStefano Zampini } 1820ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 182180677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 182280677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1823ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 182480677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 182506656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1826ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1827ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 182806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 182906656605SStefano Zampini } 183004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 183104708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 183204708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 183304708bb6SStefano Zampini } 1834ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1835ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1836ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1837ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1838ffd830a3SStefano Zampini } 183906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 184014393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 184114393ed6SStefano Zampini if (need_benign_correction) { 1842*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 184314393ed6SStefano Zampini PetscScalar *marr,*sums; 184414393ed6SStefano Zampini 184514393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 184614393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1847*df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 184814393ed6SStefano Zampini const PetscScalar *vals; 184914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 185014393ed6SStefano Zampini PetscInt n,j,nz; 185114393ed6SStefano Zampini 1852*df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1853*df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 185414393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 185514393ed6SStefano Zampini PetscInt k; 185614393ed6SStefano Zampini sums[j] = 0.; 185714393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 185814393ed6SStefano Zampini } 185914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 186014393ed6SStefano Zampini for (j=0;j<n;j++) { 186114393ed6SStefano Zampini PetscScalar val = vals[j]; 186214393ed6SStefano Zampini PetscInt k; 186314393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 186414393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 186514393ed6SStefano Zampini } 186614393ed6SStefano Zampini } 186714393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1868*df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186914393ed6SStefano Zampini } 187014393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 187114393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 187214393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 187314393ed6SStefano Zampini } 187480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 187506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 187606656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 187706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 187806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 187906656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 188006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 188106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1882d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1883019a44ceSStefano Zampini } else { 1884d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1885d16cbb6bSStefano Zampini } 18864f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1887019a44ceSStefano Zampini const PetscScalar *vals; 1888019a44ceSStefano Zampini const PetscInt *idxs; 18894f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1890019a44ceSStefano Zampini 18914f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 18924f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1893d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 18944f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 18954f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1896019a44ceSStefano Zampini } 18974f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 189816f15bc4SStefano Zampini } 189921eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1900d16cbb6bSStefano Zampini 190106656605SStefano Zampini /* coarse basis functions */ 190206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 190316f15bc4SStefano Zampini PetscScalar *y; 190416f15bc4SStefano Zampini 1905ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 190606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 190706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 190806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 190906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 191006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 191106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 191206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 191306656605SStefano Zampini 191406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19154f1b2e48SStefano Zampini PetscInt j; 19164f1b2e48SStefano Zampini 191706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 191806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 191906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19224f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 192306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 192406656605SStefano Zampini } 192506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 192606656605SStefano Zampini } 192704708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 192804708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 192906656605SStefano Zampini } 193006656605SStefano Zampini 193106656605SStefano Zampini if (n_constraints) { 193206656605SStefano Zampini Mat B; 193306656605SStefano Zampini 1934ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 193506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 193680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 193706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 193806656605SStefano Zampini if (n_vertices) { 193980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 194080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 194180677318SStefano Zampini } else { 194280677318SStefano Zampini Mat S_VCt; 194380677318SStefano Zampini 1944ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1945ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1946ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1947ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1948ffd830a3SStefano Zampini } 194980677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 195080677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 195180677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 195280677318SStefano Zampini } 195306656605SStefano Zampini } 195406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195506656605SStefano Zampini /* coarse basis functions */ 195606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 195706656605SStefano Zampini PetscScalar *y; 195806656605SStefano Zampini 1959ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 196006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 196106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 196206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 196506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 196606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19674f1b2e48SStefano Zampini PetscInt j; 19684f1b2e48SStefano Zampini 196906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 197006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 197106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19744f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 197506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 197606656605SStefano Zampini } 197706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 197806656605SStefano Zampini } 197906656605SStefano Zampini } 198080677318SStefano Zampini if (n_constraints) { 198180677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 198280677318SStefano Zampini } 19834f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 1984019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 1985019a44ceSStefano Zampini 198606656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 19873301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 1988ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 1989ffd830a3SStefano Zampini PetscScalar *marray; 199006656605SStefano Zampini 199106656605SStefano Zampini if (n_constraints) { 1992ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 199306656605SStefano Zampini 1994ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 199506656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 1996ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 199716f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 199806656605SStefano Zampini if (n_vertices) { 1999ffd830a3SStefano Zampini Mat S_VCT; 200006656605SStefano Zampini 200106656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2002ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 200316f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 200406656605SStefano Zampini } 2005ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 200606656605SStefano Zampini } 200716f15bc4SStefano Zampini if (n_vertices && n_R) { 2008ffd830a3SStefano Zampini PetscScalar *av,*marray; 2009ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2010ffd830a3SStefano Zampini PetscInt n; 2011ffd830a3SStefano Zampini PetscBool flg_row; 201206656605SStefano Zampini 2013ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2014ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2015ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2016ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2017ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2018ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2019ffd830a3SStefano Zampini PetscInt j; 2020ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2021ffd830a3SStefano Zampini } 2022ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2023ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2024ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 202506656605SStefano Zampini } 202606656605SStefano Zampini 2027ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2028ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2029ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2030ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2031ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 203206656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 203306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 203406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 203506656605SStefano Zampini } 2036ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2037ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2038ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2039ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2040ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2041ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2042ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2043ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 204406656605SStefano Zampini } 2045ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 204606656605SStefano Zampini /* coarse basis functions */ 204706656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 204806656605SStefano Zampini PetscScalar *y; 204906656605SStefano Zampini 2050ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 205106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 205206656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 205306656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205406656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205506656605SStefano Zampini if (i<n_vertices) { 205606656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 205706656605SStefano Zampini } 205806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 205906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 206006656605SStefano Zampini 206106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 206206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 206306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 206406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 206606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 206706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 206806656605SStefano Zampini } 206906656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 207006656605SStefano Zampini } 2071ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2072ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 207306656605SStefano Zampini } 2074d62866d3SStefano Zampini /* free memory */ 207588ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 207606656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 207706656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 207806656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 207906656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2080d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2081d62866d3SStefano Zampini if (n_vertices) { 2082d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2083d62866d3SStefano Zampini } 2084d62866d3SStefano Zampini if (n_constraints) { 2085d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2086d62866d3SStefano Zampini } 208788ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 208888ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 208988ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2090d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 209188ebb749SStefano Zampini Mat coarse_sub_mat; 209225084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 209388ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 209488ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 209588ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 20968bec7fa6SStefano Zampini Mat C_B,CPHI; 20978bec7fa6SStefano Zampini IS is_dummy; 20988bec7fa6SStefano Zampini Vec mones; 209988ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 210088ebb749SStefano Zampini PetscReal real_value; 210188ebb749SStefano Zampini 2102a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2103a3df083aSStefano Zampini Mat A; 2104a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2105a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2106a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2107a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2108a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2109a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2110a3df083aSStefano Zampini } else { 211188ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 211288ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 211388ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 211488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2115a3df083aSStefano Zampini } 211688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 211788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2118ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 211988ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 212088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 212188ebb749SStefano Zampini } 212288ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 212388ebb749SStefano Zampini 212425084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 21253301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 212625084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2127ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 212888ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 212988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 213088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 213188ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 213288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 213388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 213488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 213588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 213688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 213788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 213888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 213988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 214088ebb749SStefano Zampini } else { 214188ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 214288ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 214388ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 214488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 214588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 214688ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 214788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 214888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 214988ebb749SStefano Zampini } 215088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 215188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 215288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 215388ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 21544f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2155d12edf2fSStefano Zampini Mat B0_I,B0_B,B0_BPHI,B0_IPHI; 2156d12edf2fSStefano Zampini PetscScalar *data,*data2; 21574f1b2e48SStefano Zampini PetscInt j; 2158d12edf2fSStefano Zampini 21594f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 21604f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B); 21614f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_I_local,MAT_INITIAL_MATRIX,&B0_I); 2162d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2163d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2164d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2165d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 21664f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 21674f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2168d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 21694f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 21704f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 21714f1b2e48SStefano Zampini } 2172d12edf2fSStefano Zampini } 2173d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2174d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2175d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2176d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2177d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2178d12edf2fSStefano Zampini ierr = MatMatMult(B0_I,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&B0_IPHI);CHKERRQ(ierr); 2179d12edf2fSStefano Zampini ierr = MatDestroy(&B0_I);CHKERRQ(ierr); 2180d12edf2fSStefano Zampini ierr = MatNorm(B0_IPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2181d12edf2fSStefano Zampini ierr = MatDestroy(&B0_IPHI);CHKERRQ(ierr); 2182d12edf2fSStefano Zampini } 2183d12edf2fSStefano Zampini #if 0 2184d12edf2fSStefano Zampini { 2185d12edf2fSStefano Zampini PetscViewer viewer; 2186d12edf2fSStefano Zampini char filename[256]; 2187ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2188d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2189d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2190ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2191ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2192ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2193d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 2194ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2195ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 2196ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2197ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2198ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2199ffd830a3SStefano Zampini } 2200ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2201ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2202ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2203ffd830a3SStefano Zampini } 2204ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 2205ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2206ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2207ffd830a3SStefano Zampini } 2208d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2209d12edf2fSStefano Zampini } 2210d12edf2fSStefano Zampini #endif 221181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 22128bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 22131575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 221406656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 22158bec7fa6SStefano Zampini 22168bec7fa6SStefano Zampini /* check constraints */ 22174f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 22188bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 22198bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 22208bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 22218bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 22228bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 22238bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 22248bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2225bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2226ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2227bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2228bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2229bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2230bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2231bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 223288ebb749SStefano Zampini } 22338bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 22348bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 22358bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 22368bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2237d12edf2fSStefano Zampini } 223825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 223988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 224088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 224188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 224288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 224388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 224488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 224588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 224688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 224788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 224888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2249ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 225088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 225188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 225288ebb749SStefano Zampini } 225388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 225488ebb749SStefano Zampini } 22558629588bSStefano Zampini /* get back data */ 22568629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 225788ebb749SStefano Zampini PetscFunctionReturn(0); 225888ebb749SStefano Zampini } 225988ebb749SStefano Zampini 226088ebb749SStefano Zampini #undef __FUNCT__ 2261d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2262d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2263aa0d41d4SStefano Zampini { 2264d65f70fdSStefano Zampini Mat *work_mat; 2265d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2266d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2267d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2268aa0d41d4SStefano Zampini PetscErrorCode ierr; 2269aa0d41d4SStefano Zampini 2270aa0d41d4SStefano Zampini PetscFunctionBegin; 2271d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2272d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2273d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2274d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2275aa0d41d4SStefano Zampini 2276d65f70fdSStefano Zampini if (!rsorted) { 2277906d46d4SStefano Zampini const PetscInt *idxs; 2278906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2279aa0d41d4SStefano Zampini 2280d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2281d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2282d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2283d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2284aa0d41d4SStefano Zampini } 2285d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2286d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2287d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2288d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2289aa0d41d4SStefano Zampini } 2290d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2291d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2292d65f70fdSStefano Zampini } else { 2293d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2294d65f70fdSStefano Zampini isrow_s = isrow; 2295aa0d41d4SStefano Zampini } 2296906d46d4SStefano Zampini 2297d65f70fdSStefano Zampini if (!csorted) { 2298d65f70fdSStefano Zampini if (isrow == iscol) { 2299d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2300d65f70fdSStefano Zampini iscol_s = isrow_s; 2301d65f70fdSStefano Zampini } else { 2302d65f70fdSStefano Zampini const PetscInt *idxs; 2303d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2304906d46d4SStefano Zampini 2305d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2306d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2307d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2308d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2309d65f70fdSStefano Zampini } 2310d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2311d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2312d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2313d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2314d65f70fdSStefano Zampini } 2315d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2316d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2317d65f70fdSStefano Zampini } 2318d65f70fdSStefano Zampini } else { 2319d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2320d65f70fdSStefano Zampini iscol_s = iscol; 2321d65f70fdSStefano Zampini } 2322d65f70fdSStefano Zampini 2323d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2324d65f70fdSStefano Zampini 2325d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2326906d46d4SStefano Zampini Mat new_mat; 2327d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2328906d46d4SStefano Zampini 2329d65f70fdSStefano Zampini if (!rsorted) { 2330d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2331d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2332d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2333d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2334906d46d4SStefano Zampini } 2335d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2336d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2337d65f70fdSStefano Zampini } else { 2338d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2339906d46d4SStefano Zampini } 2340d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2341d65f70fdSStefano Zampini 2342d65f70fdSStefano Zampini if (!csorted) { 2343d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2344d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2345d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2346d65f70fdSStefano Zampini } else { 2347d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2348d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2349d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2350d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2351d65f70fdSStefano Zampini } 2352d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2353d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2354d65f70fdSStefano Zampini } 2355d65f70fdSStefano Zampini } else { 2356d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2357d65f70fdSStefano Zampini } 2358d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2359d65f70fdSStefano Zampini 2360d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2361d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2362d65f70fdSStefano Zampini work_mat[0] = new_mat; 2363d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2364d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2365d65f70fdSStefano Zampini } 2366d65f70fdSStefano Zampini 2367d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2368d65f70fdSStefano Zampini *B = work_mat[0]; 2369d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2370d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2371d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2372d65f70fdSStefano Zampini PetscFunctionReturn(0); 2373d65f70fdSStefano Zampini } 2374d65f70fdSStefano Zampini 2375d65f70fdSStefano Zampini #undef __FUNCT__ 23765e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 23775e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2378aa0d41d4SStefano Zampini { 2379aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 23805e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2381d65f70fdSStefano Zampini Mat new_mat; 23825e8657edSStefano Zampini IS is_local,is_global; 2383d65f70fdSStefano Zampini PetscInt local_size; 2384d65f70fdSStefano Zampini PetscBool isseqaij; 2385aa0d41d4SStefano Zampini PetscErrorCode ierr; 2386aa0d41d4SStefano Zampini 2387aa0d41d4SStefano Zampini PetscFunctionBegin; 2388aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 23895e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 23905e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2391b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2392aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2393d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2394aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2395906d46d4SStefano Zampini 2396906d46d4SStefano Zampini /* check */ 2397906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2398906d46d4SStefano Zampini Vec x,x_change; 2399906d46d4SStefano Zampini PetscReal error; 2400906d46d4SStefano Zampini 24015e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2402906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 24035e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2404e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2405e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2406d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2407e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2408e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2409906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2410906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2411906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2412906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2413906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2414906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2415906d46d4SStefano Zampini } 2416906d46d4SStefano Zampini 241722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 24189b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 241922d5777bSStefano Zampini if (isseqaij) { 24201cf9b237SStefano Zampini Mat M; 24211cf9b237SStefano Zampini 24221cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24231cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24241cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2425aa0d41d4SStefano Zampini } else { 24261cf9b237SStefano Zampini Mat work_mat,M; 24271cf9b237SStefano Zampini 2428aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 24291cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24301cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24311cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2432aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2433aa0d41d4SStefano Zampini } 24343301b35fSStefano Zampini if (matis->A->symmetric_set) { 24353301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2436e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24373301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2438e496cd5dSStefano Zampini #endif 24393301b35fSStefano Zampini } 244045a1bb75SStefano Zampini /* 244145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2442d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 244345a1bb75SStefano Zampini */ 2444d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2445aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2446aa0d41d4SStefano Zampini } 2447aa0d41d4SStefano Zampini 2448aa0d41d4SStefano Zampini #undef __FUNCT__ 2449a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 24508ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2451a64d13efSStefano Zampini { 2452a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2453a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2454d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 245553892102SStefano Zampini PetscInt *idx_R_local=NULL; 24563a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 24573a50541eSStefano Zampini PetscInt vbs,bs; 24586816873aSStefano Zampini PetscBT bitmask=NULL; 2459a64d13efSStefano Zampini PetscErrorCode ierr; 2460a64d13efSStefano Zampini 2461a64d13efSStefano Zampini PetscFunctionBegin; 2462b23d619eSStefano Zampini /* 2463b23d619eSStefano Zampini No need to setup local scatters if 2464b23d619eSStefano Zampini - primal space is unchanged 2465b23d619eSStefano Zampini AND 2466b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2467b23d619eSStefano Zampini AND 2468b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2469b23d619eSStefano Zampini */ 2470b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2471f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2472f4ddd8eeSStefano Zampini } 2473f4ddd8eeSStefano Zampini /* destroy old objects */ 2474f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2475f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2476f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2477a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2478b371cd4fSStefano Zampini n_B = pcis->n_B; 2479b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2480b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 24813a50541eSStefano Zampini 2482a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 24836816873aSStefano Zampini 248453892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2485*df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2486854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2487a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2488a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 24890e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2490a64d13efSStefano Zampini } 2491a64d13efSStefano Zampini 2492a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 24934641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 24946816873aSStefano Zampini idx_R_local[n_R++] = i; 2495a64d13efSStefano Zampini } 2496a64d13efSStefano Zampini } 2497*df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2498*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 24996816873aSStefano Zampini 2500*df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2501*df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 25026816873aSStefano Zampini } 25033a50541eSStefano Zampini 25043a50541eSStefano Zampini /* Block code */ 25053a50541eSStefano Zampini vbs = 1; 25063a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 25073a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 25083a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 25093a50541eSStefano Zampini PetscInt *vary; 2510*df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2511785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 25123a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2513d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2514d3df7717SStefano 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 */ 25150e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2516d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 25173a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 25183a50541eSStefano Zampini is_blocked = PETSC_FALSE; 25193a50541eSStefano Zampini break; 25203a50541eSStefano Zampini } 25213a50541eSStefano Zampini } 2522d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2523d3df7717SStefano Zampini } else { 2524d3df7717SStefano Zampini /* Verify directly the R set */ 2525d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2526d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2527d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2528d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2529d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2530d3df7717SStefano Zampini break; 2531d3df7717SStefano Zampini } 2532d3df7717SStefano Zampini } 2533d3df7717SStefano Zampini } 2534d3df7717SStefano Zampini } 25353a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 25363a50541eSStefano Zampini vbs = bs; 25373a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 25383a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 25393a50541eSStefano Zampini } 25403a50541eSStefano Zampini } 25413a50541eSStefano Zampini } 25423a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2543*df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2544*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 254553892102SStefano Zampini 2546*df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2547*df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 254853892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2549*df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 255053892102SStefano Zampini } else { 25513a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 255253892102SStefano Zampini } 2553a64d13efSStefano Zampini 2554a64d13efSStefano Zampini /* print some info if requested */ 2555a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2556a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2557a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 25581575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2559a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2560a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 25614f1b2e48SStefano 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); 2562a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2563a64d13efSStefano Zampini } 2564a64d13efSStefano Zampini 2565a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2566*df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 25676816873aSStefano Zampini IS is_aux1,is_aux2; 25686816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 25696816873aSStefano Zampini 25703a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2571854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2572854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2573a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 25744641a718SStefano Zampini for (i=0; i<n_D; i++) { 25754641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 25764641a718SStefano Zampini } 2577a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2578a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 25794641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 25804641a718SStefano Zampini aux_array1[j++] = i; 2581a64d13efSStefano Zampini } 2582a64d13efSStefano Zampini } 2583a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2584a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2585a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 25864641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 25874641a718SStefano Zampini aux_array2[j++] = i; 2588a64d13efSStefano Zampini } 2589a64d13efSStefano Zampini } 2590a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2591a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2592a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2593a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2594a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2595a64d13efSStefano Zampini 25968eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2597785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2598a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 25994641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 26004641a718SStefano Zampini aux_array1[j++] = i; 2601a64d13efSStefano Zampini } 2602a64d13efSStefano Zampini } 2603a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2604a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2605a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2606a64d13efSStefano Zampini } 26074641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 26083a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2609d62866d3SStefano Zampini } else { 2610*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26116816873aSStefano Zampini IS tis; 26126816873aSStefano Zampini PetscInt schur_size; 26136816873aSStefano Zampini 2614*df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 26156816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2616*df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 26176816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 26186816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 26196816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 26206816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 26216816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2622d62866d3SStefano Zampini } 2623d62866d3SStefano Zampini } 2624a64d13efSStefano Zampini PetscFunctionReturn(0); 2625a64d13efSStefano Zampini } 2626a64d13efSStefano Zampini 2627304d26faSStefano Zampini 2628304d26faSStefano Zampini #undef __FUNCT__ 2629304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2630684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2631304d26faSStefano Zampini { 2632304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2633304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2634304d26faSStefano Zampini PC pc_temp; 2635304d26faSStefano Zampini Mat A_RR; 2636f4ddd8eeSStefano Zampini MatReuse reuse; 2637304d26faSStefano Zampini PetscScalar m_one = -1.0; 2638304d26faSStefano Zampini PetscReal value; 263904708bb6SStefano Zampini PetscInt n_D,n_R; 26409577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2641304d26faSStefano Zampini PetscErrorCode ierr; 2642e604994aSStefano Zampini /* prefixes stuff */ 2643312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2644e604994aSStefano Zampini size_t len; 2645304d26faSStefano Zampini 2646304d26faSStefano Zampini PetscFunctionBegin; 2647304d26faSStefano Zampini 2648e604994aSStefano Zampini /* compute prefixes */ 2649e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2650e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2651e604994aSStefano Zampini if (!pcbddc->current_level) { 2652e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2653e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2654e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2655e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2656e604994aSStefano Zampini } else { 2657e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2658312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2659e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2660e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2661312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2662312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 266334d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 266434d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);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 ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2668e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2669e604994aSStefano Zampini } 2670e604994aSStefano Zampini 2671304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2672684f6988SStefano Zampini if (dirichlet) { 2673d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2674a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2675a3df083aSStefano Zampini Mat A_IIn; 2676a3df083aSStefano Zampini 2677a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2678a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2679a3df083aSStefano Zampini pcis->A_II = A_IIn; 2680a3df083aSStefano Zampini } 26813301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 26823301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2683964fefecSStefano Zampini } 2684ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2685964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2686304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2687304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2688304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2689304d26faSStefano Zampini /* default */ 2690304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2691e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 26929577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2693304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 26949577ea80SStefano Zampini if (issbaij) { 26959577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 26969577ea80SStefano Zampini } else { 2697304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 26989577ea80SStefano Zampini } 2699304d26faSStefano Zampini /* Allow user's customization */ 2700304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2701304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2702304d26faSStefano Zampini } 2703d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2704*df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2705*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2706d62866d3SStefano Zampini 2707*df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2708d5574798SStefano Zampini } 2709304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2710304d26faSStefano Zampini if (!n_D) { 2711304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2712304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2713304d26faSStefano Zampini } 2714304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2715304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2716304d26faSStefano Zampini /* set ksp_D into pcis data */ 2717304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2718304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2719304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2720684f6988SStefano Zampini } 2721304d26faSStefano Zampini 2722304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2723684f6988SStefano Zampini A_RR = 0; 2724684f6988SStefano Zampini if (neumann) { 2725d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 272604708bb6SStefano Zampini PetscInt ibs,mbs; 272704708bb6SStefano Zampini PetscBool issbaij; 272804708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2729f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 27308ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2731f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2732f4ddd8eeSStefano Zampini PetscInt nn_R; 273381d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2734f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2735f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2736f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2737f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2738f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2739f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2740f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2741727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2742f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2743f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2744f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2745f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2746f4ddd8eeSStefano Zampini } 2747f4ddd8eeSStefano Zampini } 2748f4ddd8eeSStefano Zampini /* last check */ 2749d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2750f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2751f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2752f4ddd8eeSStefano Zampini } 2753f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2754f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2755f4ddd8eeSStefano Zampini } 2756f4ddd8eeSStefano Zampini /* extract A_RR */ 2757af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2758af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 275904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 276004708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 276104708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 276204708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 276304708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2764af732b37SStefano Zampini } else { 276504708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 27666816873aSStefano Zampini } 276704708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 276804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 276904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 277004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 277104708bb6SStefano Zampini } else { 277204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 277304708bb6SStefano Zampini } 277404708bb6SStefano Zampini } 2775f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 27763301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27773301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 27786816873aSStefano Zampini } 2779a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2780a3df083aSStefano Zampini Mat A_RRn; 2781a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RRn);CHKERRQ(ierr); 2782a3df083aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2783a3df083aSStefano Zampini A_RR = A_RRn; 2784a3df083aSStefano Zampini } 2785f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2786304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2787304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2788304d26faSStefano Zampini /* default */ 2789304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2790e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2791304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 27929577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 27939577ea80SStefano Zampini if (issbaij) { 27949577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 27959577ea80SStefano Zampini } else { 2796304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 27979577ea80SStefano Zampini } 2798304d26faSStefano Zampini /* Allow user's customization */ 2799304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2800304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2801304d26faSStefano Zampini } 2802304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2803304d26faSStefano Zampini if (!n_R) { 2804304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2805304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2806304d26faSStefano Zampini } 2807*df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2808*df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2809*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2810d62866d3SStefano Zampini 2811*df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2812d62866d3SStefano Zampini } 2813ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2814304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2815304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2816684f6988SStefano Zampini } 28176816873aSStefano Zampini /* free Neumann problem's matrix */ 28186816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2819304d26faSStefano Zampini 2820304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 28210fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2822684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2823684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28241575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2825684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2826684f6988SStefano Zampini } 2827684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 28280fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 28290fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 28300fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 28310fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 28320fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2833304d26faSStefano Zampini /* need to be adapted? */ 2834b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2835b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2836b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2837304d26faSStefano Zampini /* print info */ 2838304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2839e604994aSStefano 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); 2840304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2841304d26faSStefano Zampini } 2842b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2843298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2844304d26faSStefano Zampini } 2845684f6988SStefano Zampini } 2846684f6988SStefano Zampini if (neumann) { /* Neumann */ 28476816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 28480fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 28490fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 28500fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 28510fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 28520fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2853304d26faSStefano Zampini /* need to be adapted? */ 2854b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2855b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2856304d26faSStefano Zampini /* print info */ 2857304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2858e604994aSStefano 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); 2859304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2860304d26faSStefano Zampini } 2861b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2862298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2863304d26faSStefano Zampini } 28640fccc4e9SStefano Zampini } 2865684f6988SStefano Zampini } 2866304d26faSStefano Zampini PetscFunctionReturn(0); 2867304d26faSStefano Zampini } 2868304d26faSStefano Zampini 2869304d26faSStefano Zampini #undef __FUNCT__ 2870ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 287180677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2872674ae819SStefano Zampini { 2873674ae819SStefano Zampini PetscErrorCode ierr; 2874674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2875be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2876674ae819SStefano Zampini 2877674ae819SStefano Zampini PetscFunctionBegin; 2878*df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 287980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 288020c7b377SStefano Zampini } 288180677318SStefano Zampini if (!pcbddc->switch_static) { 288280677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 288380677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 288480677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 288520c7b377SStefano Zampini } 2886*df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 288780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 288880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 288920c7b377SStefano Zampini } else { 2890*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2891be83ff47SStefano Zampini 2892*df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2893*df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 289420c7b377SStefano Zampini } 2895be83ff47SStefano Zampini } else { 289680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 289780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 289880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 289980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 290080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 290180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 290280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 290380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 290480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2905674ae819SStefano Zampini } 2906674ae819SStefano Zampini } 2907*df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 290880677318SStefano Zampini if (applytranspose) { 290980677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 291080677318SStefano Zampini } else { 291180677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 291280677318SStefano Zampini } 2913be83ff47SStefano Zampini } else { 2914*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2915be83ff47SStefano Zampini 2916be83ff47SStefano Zampini if (applytranspose) { 2917*df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2918be83ff47SStefano Zampini } else { 2919*df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2920be83ff47SStefano Zampini } 2921be83ff47SStefano Zampini } 292280677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 292380677318SStefano Zampini if (!pcbddc->switch_static) { 2924*df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 292580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 292680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2927be83ff47SStefano Zampini } else { 2928*df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2929be83ff47SStefano Zampini 2930*df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2931*df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2932be83ff47SStefano Zampini } 293380677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 293480677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 293580677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 293680677318SStefano Zampini } 293780677318SStefano Zampini } else { 293880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 293980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 294080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 294180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 294280677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 294380677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 294480677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 294580677318SStefano Zampini } 294680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 294780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 294880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 294980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2950674ae819SStefano Zampini } 2951674ae819SStefano Zampini PetscFunctionReturn(0); 2952674ae819SStefano Zampini } 2953674ae819SStefano Zampini 2954dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2955674ae819SStefano Zampini #undef __FUNCT__ 2956674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2957dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2958674ae819SStefano Zampini { 2959674ae819SStefano Zampini PetscErrorCode ierr; 2960674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2961674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 2962674ae819SStefano Zampini const PetscScalar zero = 0.0; 2963674ae819SStefano Zampini 2964674ae819SStefano Zampini PetscFunctionBegin; 2965dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 2966dc359a40SStefano Zampini if (applytranspose) { 2967674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 29688eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 2969dc359a40SStefano Zampini } else { 2970674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 2971674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 297215aaf578SStefano Zampini } 2973efc2fbd9SStefano Zampini 2974efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 29754f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2976efc2fbd9SStefano Zampini PetscScalar *array; 29774f1b2e48SStefano Zampini PetscInt j; 2978efc2fbd9SStefano Zampini 2979efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 29804f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 2981efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 2982efc2fbd9SStefano Zampini } 2983efc2fbd9SStefano Zampini 298412edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 298512edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 298612edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298812edc857SStefano Zampini 29899f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 299012edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 299112edc857SStefano Zampini if (pcbddc->coarse_ksp) { 299251694757SStefano Zampini Mat coarse_mat; 2993964fefecSStefano Zampini Vec rhs,sol; 299451694757SStefano Zampini MatNullSpace nullsp; 2995964fefecSStefano Zampini 2996964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 2997964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 299851694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 299951694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 300051694757SStefano Zampini if (nullsp) { 300151694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 300251694757SStefano Zampini } 300312edc857SStefano Zampini if (applytranspose) { 3004964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 300512edc857SStefano Zampini } else { 3006964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 300712edc857SStefano Zampini } 300851694757SStefano Zampini if (nullsp) { 300951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 301051694757SStefano Zampini } 301112edc857SStefano Zampini } 3012674ae819SStefano Zampini 3013674ae819SStefano Zampini /* Local solution on R nodes */ 301480677318SStefano Zampini if (pcis->n) { /* in/out pcbddc->vec1_B,pcbddc->vec1_D */ 301580677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 30169f00e9b4SStefano Zampini } 3017674ae819SStefano Zampini 30189f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 30199f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 302012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3021674ae819SStefano Zampini 3022674ae819SStefano Zampini /* Sum contributions from two levels */ 3023dc359a40SStefano Zampini if (applytranspose) { 3024dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3025dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3026dc359a40SStefano Zampini } else { 3027674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 30288eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3029dc359a40SStefano Zampini } 3030efc2fbd9SStefano Zampini /* store p0 */ 30314f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3032efc2fbd9SStefano Zampini PetscScalar *array; 30334f1b2e48SStefano Zampini PetscInt j; 3034efc2fbd9SStefano Zampini 3035efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30364f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3037efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3038efc2fbd9SStefano Zampini } 3039674ae819SStefano Zampini PetscFunctionReturn(0); 3040674ae819SStefano Zampini } 3041674ae819SStefano Zampini 3042674ae819SStefano Zampini #undef __FUNCT__ 3043674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 304412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3045674ae819SStefano Zampini { 3046674ae819SStefano Zampini PetscErrorCode ierr; 3047674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 304858da7f69SStefano Zampini PetscScalar *array; 304912edc857SStefano Zampini Vec from,to; 3050674ae819SStefano Zampini 3051674ae819SStefano Zampini PetscFunctionBegin; 305212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 305312edc857SStefano Zampini from = pcbddc->coarse_vec; 305412edc857SStefano Zampini to = pcbddc->vec1_P; 305512edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 305612edc857SStefano Zampini Vec tvec; 305758da7f69SStefano Zampini 305858da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 305958da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 306012edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 306158da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 306258da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 306358da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 306412edc857SStefano Zampini } 306512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 306612edc857SStefano Zampini from = pcbddc->vec1_P; 306712edc857SStefano Zampini to = pcbddc->coarse_vec; 306812edc857SStefano Zampini } 306912edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3070674ae819SStefano Zampini PetscFunctionReturn(0); 3071674ae819SStefano Zampini } 3072674ae819SStefano Zampini 3073674ae819SStefano Zampini #undef __FUNCT__ 3074674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 307512edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3076674ae819SStefano Zampini { 3077674ae819SStefano Zampini PetscErrorCode ierr; 3078674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 307958da7f69SStefano Zampini PetscScalar *array; 308012edc857SStefano Zampini Vec from,to; 3081674ae819SStefano Zampini 3082674ae819SStefano Zampini PetscFunctionBegin; 308312edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 308412edc857SStefano Zampini from = pcbddc->coarse_vec; 308512edc857SStefano Zampini to = pcbddc->vec1_P; 308612edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 308712edc857SStefano Zampini from = pcbddc->vec1_P; 308812edc857SStefano Zampini to = pcbddc->coarse_vec; 308912edc857SStefano Zampini } 309012edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 309112edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 309212edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 309312edc857SStefano Zampini Vec tvec; 309458da7f69SStefano Zampini 309512edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 309658da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 309758da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 309858da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 309958da7f69SStefano Zampini } 310058da7f69SStefano Zampini } else { 310158da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 310258da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 310312edc857SStefano Zampini } 310412edc857SStefano Zampini } 3105674ae819SStefano Zampini PetscFunctionReturn(0); 3106674ae819SStefano Zampini } 3107674ae819SStefano Zampini 3108984c4197SStefano Zampini /* uncomment for testing purposes */ 3109984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3110674ae819SStefano Zampini #undef __FUNCT__ 3111674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3112674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3113674ae819SStefano Zampini { 3114674ae819SStefano Zampini PetscErrorCode ierr; 3115674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3116674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3117674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3118984c4197SStefano Zampini /* one and zero */ 3119984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3120984c4197SStefano Zampini /* space to store constraints and their local indices */ 31219162d606SStefano Zampini PetscScalar *constraints_data; 31229162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 31239162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 31249162d606SStefano Zampini PetscInt *constraints_n; 3125984c4197SStefano Zampini /* iterators */ 3126b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3127984c4197SStefano Zampini /* BLAS integers */ 3128e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3129e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3130c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3131727cdba6SStefano Zampini /* reuse */ 31320e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 31330e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3134984c4197SStefano Zampini /* change of basis */ 3135b3d85658SStefano Zampini PetscBool qr_needed; 31369162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3137984c4197SStefano Zampini /* auxiliary stuff */ 313864efe560SStefano Zampini PetscInt *nnz,*is_indices; 31398a0068c3SStefano Zampini PetscInt ncc; 3140984c4197SStefano Zampini /* some quantities */ 314145a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3142a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3143984c4197SStefano Zampini 3144674ae819SStefano Zampini PetscFunctionBegin; 31458e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 31468e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 31478e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3148088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3149088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 31500e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 31510e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 31520e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 31530e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 31540e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3155088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3156cf5a6209SStefano Zampini 3157cf5a6209SStefano Zampini /* print some info */ 3158cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3159cf5a6209SStefano Zampini IS vertices; 3160cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3161cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3162cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3163cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 31641575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3165cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3166cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3167fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3168fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3169cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 31701575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3171cf5a6209SStefano Zampini } 3172cf5a6209SStefano Zampini 3173cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 31749162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3175cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3176cf5a6209SStefano Zampini const Vec *nearnullvecs; 3177cf5a6209SStefano Zampini Vec *localnearnullsp; 3178cf5a6209SStefano Zampini PetscScalar *array; 3179cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3180cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3181674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3182b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3183674ae819SStefano Zampini PetscScalar *work; 3184674ae819SStefano Zampini PetscReal *singular_vals; 3185674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3186674ae819SStefano Zampini PetscReal *rwork; 3187674ae819SStefano Zampini #endif 3188674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3189674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3190674ae819SStefano Zampini #else 3191964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3192964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3193674ae819SStefano Zampini #endif 3194674ae819SStefano Zampini 3195674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3196d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3197d06fc5fdSStefano Zampini /* free unneeded index sets */ 3198d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3199d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3200674ae819SStefano Zampini } 3201d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3202d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3203d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3204d06fc5fdSStefano Zampini } 3205d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3206d06fc5fdSStefano Zampini n_ISForEdges = 0; 3207d06fc5fdSStefano Zampini } 3208d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3209d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3210d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3211d06fc5fdSStefano Zampini } 3212d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3213d06fc5fdSStefano Zampini n_ISForFaces = 0; 3214d06fc5fdSStefano Zampini } 321570022509SStefano Zampini 321670022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 321770022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 321870022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 321970022509SStefano Zampini if (pcbddc->NullSpace) { 322070022509SStefano Zampini PetscBool tbool[2],gbool[2]; 322170022509SStefano Zampini 322270022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3223b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3224d06fc5fdSStefano Zampini if (!ISForEdges) { 3225d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3226d06fc5fdSStefano Zampini } 3227b8ffe317SStefano Zampini } 3228d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3229d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3230d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3231d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3232d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 323398a51de6SStefano Zampini } 323470022509SStefano Zampini #endif 323508122e43SStefano Zampini 3236674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3237674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3238674ae819SStefano Zampini if (nearnullsp) { 3239674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3240f4ddd8eeSStefano Zampini /* remove any stored info */ 3241f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3242f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3243f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3244f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3245f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3246473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3247f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3248f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3249f4ddd8eeSStefano Zampini } 3250984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3251984c4197SStefano Zampini nnsp_size = 0; 3252674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3253674ae819SStefano Zampini } 3254984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3255984c4197SStefano Zampini max_constraints = nnsp_size; 3256984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3257984c4197SStefano Zampini 3258674ae819SStefano Zampini /* 3259674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 32609162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 32619162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 32629162d606SStefano Zampini There can be multiple constraints per connected component 3263674ae819SStefano Zampini */ 3264674ae819SStefano Zampini n_vertices = 0; 3265674ae819SStefano Zampini if (ISForVertices) { 3266674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3267674ae819SStefano Zampini } 32689162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 32699162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 32709162d606SStefano Zampini 32719162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 32729162d606SStefano Zampini total_counts *= max_constraints; 3273674ae819SStefano Zampini total_counts += n_vertices; 32744641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 32759162d606SStefano Zampini 3276674ae819SStefano Zampini total_counts = 0; 3277674ae819SStefano Zampini max_size_of_constraint = 0; 3278674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 32799162d606SStefano Zampini IS used_is; 3280674ae819SStefano Zampini if (i<n_ISForEdges) { 32819162d606SStefano Zampini used_is = ISForEdges[i]; 3282674ae819SStefano Zampini } else { 32839162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3284674ae819SStefano Zampini } 32859162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3286674ae819SStefano Zampini total_counts += j; 3287674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3288674ae819SStefano Zampini } 32899162d606SStefano 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); 32909162d606SStefano Zampini 3291984c4197SStefano Zampini /* get local part of global near null space vectors */ 3292785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3293984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3294984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3295e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3296e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3297984c4197SStefano Zampini } 3298674ae819SStefano Zampini 3299242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3300242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3301a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3302242a89d7SStefano Zampini 3303984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3304a773dcb8SStefano Zampini if (!skip_lapack) { 3305674ae819SStefano Zampini PetscScalar temp_work; 3306911cabfeSStefano Zampini 3307674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3308984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3309785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3310785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3311785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3312674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3313785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3314674ae819SStefano Zampini #endif 3315674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3316c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3317c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3318674ae819SStefano Zampini lwork = -1; 3319674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3320674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3321c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3322674ae819SStefano Zampini #else 3323c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3324674ae819SStefano Zampini #endif 3325674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3326984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3327674ae819SStefano Zampini #else /* on missing GESVD */ 3328674ae819SStefano Zampini /* SVD */ 3329674ae819SStefano Zampini PetscInt max_n,min_n; 3330674ae819SStefano Zampini max_n = max_size_of_constraint; 3331984c4197SStefano Zampini min_n = max_constraints; 3332984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3333674ae819SStefano Zampini min_n = max_size_of_constraint; 3334984c4197SStefano Zampini max_n = max_constraints; 3335674ae819SStefano Zampini } 3336785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3337674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3338785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3339674ae819SStefano Zampini #endif 3340674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3341674ae819SStefano Zampini lwork = -1; 3342e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3343e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3344b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3345674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3346674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 33479162d606SStefano 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)); 3348674ae819SStefano Zampini #else 33499162d606SStefano 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)); 3350674ae819SStefano Zampini #endif 3351674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3352984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3353984c4197SStefano Zampini #endif /* on missing GESVD */ 3354674ae819SStefano Zampini /* Allocate optimal workspace */ 3355674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3356854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3357674ae819SStefano Zampini } 3358674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3359674ae819SStefano Zampini total_counts = 0; 33609162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 33619162d606SStefano Zampini constraints_data_ptr[0] = 0; 3362674ae819SStefano Zampini /* vertices */ 33639162d606SStefano Zampini if (n_vertices) { 3364674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 33659162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3366674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 33679162d606SStefano Zampini constraints_n[total_counts] = 1; 33689162d606SStefano Zampini constraints_data[total_counts] = 1.0; 33699162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 33709162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3371674ae819SStefano Zampini total_counts++; 3372674ae819SStefano Zampini } 3373674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3374674ae819SStefano Zampini n_vertices = total_counts; 3375674ae819SStefano Zampini } 3376984c4197SStefano Zampini 3377674ae819SStefano Zampini /* edges and faces */ 33789162d606SStefano Zampini total_counts_cc = total_counts; 3379911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 33809162d606SStefano Zampini IS used_is; 33819162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 33829162d606SStefano Zampini 3383911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 33849162d606SStefano Zampini used_is = ISForEdges[ncc]; 3385984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3386674ae819SStefano Zampini } else { 33879162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3388984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3389674ae819SStefano Zampini } 3390674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 33919162d606SStefano Zampini 33929162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 33939162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3394984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3395984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3396674ae819SStefano Zampini if (nnsp_has_cnst) { 33975b08dc53SStefano Zampini PetscScalar quad_value; 33989162d606SStefano Zampini 33999162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34009162d606SStefano Zampini idxs_copied = PETSC_TRUE; 34019162d606SStefano Zampini 3402a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3403674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3404a773dcb8SStefano Zampini } else { 3405a773dcb8SStefano Zampini quad_value = 1.0; 3406a773dcb8SStefano Zampini } 3407674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34089162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3409674ae819SStefano Zampini } 34109162d606SStefano Zampini temp_constraints++; 3411674ae819SStefano Zampini total_counts++; 3412674ae819SStefano Zampini } 3413674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3414984c4197SStefano Zampini PetscReal real_value; 34159162d606SStefano Zampini PetscScalar *ptr_to_data; 34169162d606SStefano Zampini 3417984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 34189162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3419674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34209162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3421674ae819SStefano Zampini } 3422984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3423984c4197SStefano Zampini /* check if array is null on the connected component */ 3424e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34259162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 34265b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3427674ae819SStefano Zampini temp_constraints++; 3428674ae819SStefano Zampini total_counts++; 34299162d606SStefano Zampini if (!idxs_copied) { 34309162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34319162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3432674ae819SStefano Zampini } 3433674ae819SStefano Zampini } 34349162d606SStefano Zampini } 34359162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 343645a1bb75SStefano Zampini valid_constraints = temp_constraints; 3437eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3438a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 34399162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 34409162d606SStefano Zampini 34419162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3442a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34439162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3444a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 34459162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3446a773dcb8SStefano Zampini } else { /* perform SVD */ 3447984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 34489162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3449674ae819SStefano Zampini 3450674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3451984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3452984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3453984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3454984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3455984c4197SStefano Zampini from that computed using LAPACKgesvd 3456984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3457984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3458984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3459674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3460e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3461984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3462674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3463674ae819SStefano Zampini for (k=0;k<j+1;k++) { 34649162d606SStefano 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)); 3465674ae819SStefano Zampini } 3466674ae819SStefano Zampini } 3467e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3468e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3469e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3470674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3471c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3472674ae819SStefano Zampini #else 3473c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3474674ae819SStefano Zampini #endif 3475674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3476984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3477984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3478674ae819SStefano Zampini j = 0; 3479984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3480674ae819SStefano Zampini total_counts = total_counts-j; 348145a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3482e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3483c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3484c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3485c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3486c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3487c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3488c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3489674ae819SStefano Zampini if (j<temp_constraints) { 3490984c4197SStefano Zampini PetscInt ii; 3491984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3492674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 34939162d606SStefano 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)); 3494674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3495984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3496674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 34979162d606SStefano 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]; 3498674ae819SStefano Zampini } 3499674ae819SStefano Zampini } 3500674ae819SStefano Zampini } 3501674ae819SStefano Zampini #else /* on missing GESVD */ 3502e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3503e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3504b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3505674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3506674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35079162d606SStefano 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)); 3508674ae819SStefano Zampini #else 35099162d606SStefano 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)); 3510674ae819SStefano Zampini #endif 3511984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3512674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3513984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3514e310c8b4SStefano Zampini k = temp_constraints; 3515e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3516674ae819SStefano Zampini j = 0; 3517e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 351845a1bb75SStefano Zampini valid_constraints = k-j; 3519911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3520984c4197SStefano Zampini #endif /* on missing GESVD */ 3521674ae819SStefano Zampini } 3522a773dcb8SStefano Zampini } 35239162d606SStefano Zampini /* update pointers information */ 35249162d606SStefano Zampini if (valid_constraints) { 35259162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 35269162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 35279162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 35289162d606SStefano Zampini /* set change_of_basis flag */ 352945a1bb75SStefano Zampini if (boolforchange) { 3530b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 35319162d606SStefano Zampini } 3532b3d85658SStefano Zampini total_counts_cc++; 353345a1bb75SStefano Zampini } 353445a1bb75SStefano Zampini } 3535984c4197SStefano Zampini /* free workspace */ 35368f1c130eSStefano Zampini if (!skip_lapack) { 3537984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3538984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3539984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3540984c4197SStefano Zampini #endif 3541984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3542984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3543984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3544984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3545984c4197SStefano Zampini #endif 3546984c4197SStefano Zampini } 3547984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3548984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3549984c4197SStefano Zampini } 3550984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3551cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3552cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3553cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3554cf5a6209SStefano Zampini } 3555cf5a6209SStefano Zampini if (n_ISForFaces) { 3556cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3557cf5a6209SStefano Zampini } 3558cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3559cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3560cf5a6209SStefano Zampini } 3561cf5a6209SStefano Zampini if (n_ISForEdges) { 3562cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3563cf5a6209SStefano Zampini } 3564cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 356508122e43SStefano Zampini } else { 356608122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3567984c4197SStefano Zampini 356808122e43SStefano Zampini total_counts = 0; 356908122e43SStefano Zampini n_vertices = 0; 3570d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3571d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 357208122e43SStefano Zampini } 357308122e43SStefano Zampini max_constraints = 0; 35749162d606SStefano Zampini total_counts_cc = 0; 357508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 357608122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 35779162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 357808122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 357908122e43SStefano Zampini } 35809162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 35819162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 35829162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 35839162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 358474d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 35859162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 35869162d606SStefano Zampini total_counts_cc = 0; 35879162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 35889162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 35899162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 359008122e43SStefano Zampini } 359108122e43SStefano Zampini } 35929162d606SStefano Zampini #if 0 35939162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 35949162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 35959162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 35969162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 35979162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 35989162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 35999162d606SStefano Zampini } 36009162d606SStefano Zampini printf("\n"); 36019162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 36029162d606SStefano Zampini } 36031b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 36048bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 36051b968477SStefano Zampini } 36061b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 36078bec7fa6SStefano 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]); 36081b968477SStefano Zampini } 360908122e43SStefano Zampini #endif 361008122e43SStefano Zampini 36118bec7fa6SStefano Zampini max_size_of_constraint = 0; 36129162d606SStefano 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]); 36139162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 361408122e43SStefano Zampini /* Change of basis */ 3615b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 361608122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 361708122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 361808122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3619b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 362008122e43SStefano Zampini } 362108122e43SStefano Zampini } 362208122e43SStefano Zampini } 362308122e43SStefano Zampini } 3624984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 36254f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 362608122e43SStefano Zampini 36279162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 36289162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 36299162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 36309162d606SStefano 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); 363108122e43SStefano Zampini } 3632674ae819SStefano Zampini 3633674ae819SStefano Zampini /* Create constraint matrix */ 3634674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 363516f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3636984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3637984c4197SStefano Zampini 3638984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3639a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3640a717540cSStefano Zampini qr_needed = PETSC_FALSE; 364174d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3642984c4197SStefano Zampini total_primal_vertices=0; 3643b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 36449162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36459162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3646984c4197SStefano Zampini if (size_of_constraint == 1) { 36479162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3648b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 364964efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 36509162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 36519162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3652a717540cSStefano Zampini } 3653b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 365474d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3655a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3656a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3657a717540cSStefano Zampini } 3658fa434743SStefano Zampini } else { 3659b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3660fa434743SStefano Zampini } 3661a717540cSStefano Zampini } 3662b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3663b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3664674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 366570022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3666b3d85658SStefano Zampini 36674f1b2e48SStefano 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); 36680e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 36690e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3670984c4197SStefano Zampini 3671984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 367274d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3673785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3674984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 367574d5cdf7SStefano Zampini 3676984c4197SStefano Zampini j = total_primal_vertices; 367774d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3678b3d85658SStefano Zampini cum = total_primal_vertices; 36799162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 36804641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3681b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3682b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3683b3d85658SStefano Zampini cum++; 36849162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 368574d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 368674d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 368774d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 368874d5cdf7SStefano Zampini } 36899162d606SStefano Zampini j += constraints_n[i]; 3690674ae819SStefano Zampini } 3691674ae819SStefano Zampini } 3692674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3693674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3694088faed8SStefano Zampini 3695674ae819SStefano Zampini /* set values in constraint matrix */ 3696984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 36970e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3698674ae819SStefano Zampini } 3699984c4197SStefano Zampini total_counts = total_primal_vertices; 37009162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37014641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 37029162d606SStefano Zampini PetscInt *cols; 37039162d606SStefano Zampini 37049162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 37059162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 37069162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37079162d606SStefano Zampini PetscInt row = total_counts+k; 37089162d606SStefano Zampini PetscScalar *vals; 37099162d606SStefano Zampini 37109162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 37119162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 37129162d606SStefano Zampini } 37139162d606SStefano Zampini total_counts += constraints_n[i]; 3714674ae819SStefano Zampini } 3715674ae819SStefano Zampini } 3716674ae819SStefano Zampini /* assembling */ 3717674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3718674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3719088faed8SStefano Zampini 3720984c4197SStefano Zampini /* 372145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3722984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3723984c4197SStefano Zampini */ 3724674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3725674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3726026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3727984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3728984c4197SStefano Zampini /* working stuff for GEQRF */ 372981d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3730984c4197SStefano Zampini PetscBLASInt lqr_work; 3731984c4197SStefano Zampini /* working stuff for UNGQR */ 3732984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3733984c4197SStefano Zampini PetscBLASInt lgqr_work; 3734984c4197SStefano Zampini /* working stuff for TRTRS */ 3735984c4197SStefano Zampini PetscScalar *trs_rhs; 37363f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3737984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3738984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3739984c4197SStefano Zampini PetscScalar *start_vals; 3740984c4197SStefano Zampini /* working stuff for values insertion */ 37414641a718SStefano Zampini PetscBT is_primal; 374264efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3743906d46d4SStefano Zampini /* matrix sizes */ 3744906d46d4SStefano Zampini PetscInt global_size,local_size; 3745906d46d4SStefano Zampini /* temporary change of basis */ 3746906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3747cf5a6209SStefano Zampini /* extra space for debugging */ 3748cf5a6209SStefano Zampini PetscScalar *dbg_work; 3749984c4197SStefano Zampini 3750906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3751906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 375216f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3753bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3754906d46d4SStefano Zampini /* nonzeros for local mat */ 3755bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3756bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 37579162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3758a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 37599162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3760a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 37619162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3762a717540cSStefano Zampini } else { 37639162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 37649162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3765a717540cSStefano Zampini } 3766a717540cSStefano Zampini } 3767a717540cSStefano Zampini } 3768906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3769bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3770a717540cSStefano Zampini /* Set initial identity in the matrix */ 3771bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3772906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3773a717540cSStefano Zampini } 3774a717540cSStefano Zampini 3775a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3776a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3777a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3778a717540cSStefano Zampini } 3779a717540cSStefano Zampini 3780a717540cSStefano Zampini 3781a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3782a717540cSStefano Zampini /* 3783a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3784a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3785a717540cSStefano Zampini 3786a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3787a717540cSStefano Zampini 3788a6b551f4SStefano 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) 3789a6b551f4SStefano Zampini 3790a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3791a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3792a717540cSStefano Zampini | ... | 3793a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3794a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3795a717540cSStefano Zampini 3796a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3797a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3798a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3799a6b551f4SStefano Zampini 3800a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3801a717540cSStefano Zampini */ 3802a717540cSStefano Zampini if (qr_needed) { 3803984c4197SStefano Zampini /* space to store Q */ 3804854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3805984c4197SStefano Zampini /* first we issue queries for optimal work */ 38063f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38073f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 38083f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3809984c4197SStefano Zampini lqr_work = -1; 38103f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3811984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3812984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3813785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3814984c4197SStefano Zampini lgqr_work = -1; 38153f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38163f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 38173f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 38183f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38193f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 38203f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3821984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3822984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3823785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3824984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3825785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3826984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3827785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3828a717540cSStefano Zampini /* allocating workspace for check */ 3829a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3830cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3831a717540cSStefano Zampini } 3832a717540cSStefano Zampini } 3833984c4197SStefano Zampini /* array to store whether a node is primal or not */ 38344641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3835473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 38360e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 383739e2fb2aSStefano Zampini if (i != total_primal_vertices) { 383839e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 38394641a718SStefano Zampini } 384039e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 384139e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 384239e2fb2aSStefano Zampini } 384339e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3844984c4197SStefano Zampini 3845a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 38469162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 38479162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 38484641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3849984c4197SStefano Zampini /* get constraint info */ 38509162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3851984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3852984c4197SStefano Zampini 3853984c4197SStefano Zampini if (pcbddc->dbg_flag) { 38549162d606SStefano 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); 3855674ae819SStefano Zampini } 3856984c4197SStefano Zampini 3857fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3858a717540cSStefano Zampini 3859a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3860a717540cSStefano Zampini if (pcbddc->dbg_flag) { 38619162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3862a717540cSStefano Zampini } 3863984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 38649162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3865984c4197SStefano Zampini 3866984c4197SStefano Zampini /* compute QR decomposition of constraints */ 38673f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 38683f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 38693f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3870674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 38713f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3872984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3873674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3874984c4197SStefano Zampini 3875984c4197SStefano Zampini /* explictly compute R^-T */ 3876984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3877984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 38783f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 38793f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 38803f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38813f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3882984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 38833f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3884984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3885984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3886984c4197SStefano Zampini 3887a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 38883f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 38893f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 38903f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 38913f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3892984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 38933f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3894984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3895984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3896984c4197SStefano Zampini 3897984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3898984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3899984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 39003f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39013f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39023f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39033f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39043f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 39053f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3906984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39079162d606SStefano 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)); 3908984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39099162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3910984c4197SStefano Zampini 3911984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 39129162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3913984c4197SStefano Zampini /* insert cols for primal dofs */ 3914984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3915984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 39169162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3917906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3918984c4197SStefano Zampini } 3919984c4197SStefano Zampini /* insert cols for dual dofs */ 3920984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 39219162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3922984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 39239162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3924906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3925984c4197SStefano Zampini j++; 3926674ae819SStefano Zampini } 3927674ae819SStefano Zampini } 3928984c4197SStefano Zampini 3929984c4197SStefano Zampini /* check change of basis */ 3930984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3931984c4197SStefano Zampini PetscInt ii,jj; 3932984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3933c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3934c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3935c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3936c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3937c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3938c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3939984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3940cf5a6209SStefano 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)); 3941984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3942984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3943984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3944cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3945cf5a6209SStefano 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; 3946674ae819SStefano Zampini } 3947674ae819SStefano Zampini } 3948984c4197SStefano Zampini if (!valid_qr) { 394922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3950984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3951984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3952cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 3953cf5a6209SStefano 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])); 3954674ae819SStefano Zampini } 3955cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 3956cf5a6209SStefano 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])); 3957984c4197SStefano Zampini } 3958984c4197SStefano Zampini } 3959984c4197SStefano Zampini } 3960674ae819SStefano Zampini } else { 396122d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 3962674ae819SStefano Zampini } 3963674ae819SStefano Zampini } 3964a717540cSStefano Zampini } else { /* simple transformation block */ 3965a717540cSStefano Zampini PetscInt row,col; 3966a6b551f4SStefano Zampini PetscScalar val,norm; 3967a6b551f4SStefano Zampini 3968a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39699162d606SStefano 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)); 3970a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 39719162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 39729162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3973bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 39749162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 3975906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 39769162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 3977a717540cSStefano Zampini } else { 3978a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 39799162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3980a717540cSStefano Zampini if (row != col) { 39819162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 3982a717540cSStefano Zampini } else { 39839162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 3984a717540cSStefano Zampini } 3985906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 3986a717540cSStefano Zampini } 3987a717540cSStefano Zampini } 3988a717540cSStefano Zampini } 398998a51de6SStefano Zampini if (pcbddc->dbg_flag) { 399022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 3991a717540cSStefano Zampini } 3992674ae819SStefano Zampini } 3993984c4197SStefano Zampini } else { 3994984c4197SStefano Zampini if (pcbddc->dbg_flag) { 39959162d606SStefano 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); 3996674ae819SStefano Zampini } 3997674ae819SStefano Zampini } 3998674ae819SStefano Zampini } 3999a717540cSStefano Zampini 4000a717540cSStefano Zampini /* free workspace */ 4001a717540cSStefano Zampini if (qr_needed) { 4002984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4003cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4004984c4197SStefano Zampini } 4005984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4006984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4007984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4008984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4009984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4010674ae819SStefano Zampini } 4011a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4012906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4013906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4014906d46d4SStefano Zampini 4015906d46d4SStefano Zampini /* assembling of global change of variable */ 4016bbb9e6c6SStefano Zampini { 4017bbb9e6c6SStefano Zampini Mat tmat; 401816f15bc4SStefano Zampini PetscInt bs; 401916f15bc4SStefano Zampini 4020906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4021906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4022bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4023bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4024bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4025bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 402616f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 402716f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4028906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4029bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4030bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4031bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4032bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4033bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4034e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4035e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4036bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4037bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 4038906d46d4SStefano Zampini } 4039906d46d4SStefano Zampini /* check */ 4040906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4041906d46d4SStefano Zampini PetscReal error; 4042906d46d4SStefano Zampini Vec x,x_change; 4043906d46d4SStefano Zampini 4044906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4045906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4046906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4047906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4048e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4049e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4050bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4051e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4052e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4053906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4054906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4055906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4056906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4057bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4058906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4059906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4060906d46d4SStefano Zampini } 4061b96c3477SStefano Zampini 4062b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4063b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4064b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4065b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4066ac632422SStefano Zampini Mat S_new,tmat; 4067b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 4068b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 4069b087196eSStefano Zampini const PetscScalar *array; 4070b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 4071b087196eSStefano Zampini PetscInt i,n_V; 4072bbb9e6c6SStefano Zampini 4073bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 40746816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4075b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4076b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4077b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4078b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4079bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4080b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4081ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4082b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4083ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4084b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4085b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4086b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4087b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4088b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4089b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4090b087196eSStefano Zampini PetscScalar val; 4091b087196eSStefano Zampini PetscInt idx; 4092b087196eSStefano Zampini 4093b087196eSStefano Zampini idx = idxs_V[i]; 4094b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4095b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4096b087196eSStefano Zampini } 4097b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4098b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4099ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4100ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4101ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4102ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4103b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4104ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4105b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4106ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4107ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4108ac632422SStefano Zampini } 4109b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4110b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4111b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4112b96c3477SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4113b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 4114b96c3477SStefano Zampini } 4115b96c3477SStefano Zampini } 4116906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 4117906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4118b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4119b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4120b9b85e73SStefano Zampini } 4121906d46d4SStefano Zampini 4122906d46d4SStefano Zampini /* set up change of basis context */ 4123906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4124906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4125906d46d4SStefano Zampini 4126906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4127906d46d4SStefano Zampini PetscInt global_size,local_size; 4128906d46d4SStefano Zampini 4129906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4130906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4131906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4132906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4133906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4134906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4135906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4136906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4137906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4138906d46d4SStefano Zampini } else { 4139906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4140906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4141906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4142906d46d4SStefano Zampini } 4143906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4144906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4145906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4146906d46d4SStefano Zampini } else { 4147906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4148906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4149906d46d4SStefano Zampini } 4150906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4151906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4152906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4153906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4154a8661815SStefano Zampini } else { 4155a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4156b9b85e73SStefano Zampini } 4157a717540cSStefano Zampini 41584f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 41594f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 41604f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 41614f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4162019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4163019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4164019a44ceSStefano Zampini pcbddc->local_primal_size++; 4165019a44ceSStefano Zampini } 4166019a44ceSStefano Zampini 4167019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4168727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4169727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 41709f47a83aSStefano 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); 4171c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 41720e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 41739f47a83aSStefano 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); 4174727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4175727cdba6SStefano Zampini } 41760e6343abSStefano Zampini } 41770e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4178727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4179727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4180727cdba6SStefano Zampini 4181a717540cSStefano Zampini /* flush dbg viewer */ 4182b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4183b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4184b8ffe317SStefano Zampini } 4185a717540cSStefano Zampini 4186e310c8b4SStefano Zampini /* free workspace */ 4187a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 41884641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 418908122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 41909162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 41919162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 419208122e43SStefano Zampini } else { 41939162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 41949162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 41959162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 419608122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 419708122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 41989162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 41999162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 420008122e43SStefano Zampini } 4201674ae819SStefano Zampini PetscFunctionReturn(0); 4202674ae819SStefano Zampini } 4203674ae819SStefano Zampini 4204674ae819SStefano Zampini #undef __FUNCT__ 4205674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4206674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4207674ae819SStefano Zampini { 4208674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4209674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4210674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 42117fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 4212674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4213674ae819SStefano Zampini 4214674ae819SStefano Zampini PetscFunctionBegin; 42158e61c736SStefano Zampini /* Reset previously computed graph */ 42168e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4217674ae819SStefano Zampini /* Init local Graph struct */ 42187fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 42193bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4220674ae819SStefano Zampini 4221575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 42225099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 42235099eff2SStefano 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); 4224575ad6abSStefano Zampini } 42259577ea80SStefano Zampini 4226674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 42274d379d7bSStefano Zampini if (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) { 42284d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 42294d379d7bSStefano Zampini PetscInt nvtxs; 4230e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4231674ae819SStefano Zampini 42324d379d7bSStefano Zampini if (pcbddc->use_local_adj) { 42332fffb893SStefano Zampini 42342fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42352fffb893SStefano Zampini if (flg_row) { 42364d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4237b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 42382fffb893SStefano Zampini } 42392fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42409b28b941SStefano 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 */ 42414d379d7bSStefano Zampini IS is_dummy; 42424d379d7bSStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 42434d379d7bSStefano Zampini PetscInt j,sum; 42444d379d7bSStefano Zampini PetscInt *cxadj,*cadjncy; 42454d379d7bSStefano Zampini const PetscInt *idxs; 42464d379d7bSStefano Zampini PCBDDCGraph graph; 42474d379d7bSStefano Zampini PetscBT is_on_boundary; 42484d379d7bSStefano Zampini 42494d379d7bSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcis->n,0,1,&is_dummy);CHKERRQ(ierr); 42504d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 42514d379d7bSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 42524d379d7bSStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 42537fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,pcis->n);CHKERRQ(ierr); 42544d379d7bSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 4255e496cd5dSStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4256e496cd5dSStefano Zampini if (flg_row) { 42574d379d7bSStefano Zampini graph->xadj = xadj; 42584d379d7bSStefano Zampini graph->adjncy = adjncy; 4259e496cd5dSStefano Zampini } 42604d379d7bSStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 42614d379d7bSStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 4262e496cd5dSStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 42634d379d7bSStefano Zampini 42644d379d7bSStefano Zampini if (pcbddc->dbg_flag) { 42659b28b941SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] Found %d subdomains (local size %d)\n",PetscGlobalRank,graph->ncc,pcis->n);CHKERRQ(ierr); 42664d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 42674d379d7bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"[%d] %d cc size %d\n",PetscGlobalRank,i,graph->cptr[i+1]-graph->cptr[i]);CHKERRQ(ierr); 42684d379d7bSStefano Zampini } 42694d379d7bSStefano Zampini } 42704d379d7bSStefano Zampini 4271e496cd5dSStefano Zampini ierr = PetscBTCreate(pcis->n,&is_on_boundary);CHKERRQ(ierr); 42724d379d7bSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 42734d379d7bSStefano Zampini for (i=0;i<pcis->n_B;i++) { 42744d379d7bSStefano Zampini ierr = PetscBTSet(is_on_boundary,idxs[i]);CHKERRQ(ierr); 42754d379d7bSStefano Zampini } 42764d379d7bSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 42774d379d7bSStefano Zampini 4278e496cd5dSStefano Zampini ierr = PetscCalloc1(pcis->n+1,&cxadj);CHKERRQ(ierr); 42794d379d7bSStefano Zampini sum = 0; 42804d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 42814d379d7bSStefano Zampini PetscInt sizecc = 0; 42824d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 42834d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 42844d379d7bSStefano Zampini sizecc++; 42854d379d7bSStefano Zampini } 42864d379d7bSStefano Zampini } 42874d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 42884d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 42894d379d7bSStefano Zampini cxadj[graph->queue[j]] = sizecc; 42904d379d7bSStefano Zampini } 42914d379d7bSStefano Zampini } 42924d379d7bSStefano Zampini sum += sizecc*sizecc; 42934d379d7bSStefano Zampini } 42944d379d7bSStefano Zampini ierr = PetscMalloc1(sum,&cadjncy);CHKERRQ(ierr); 42954d379d7bSStefano Zampini sum = 0; 4296e496cd5dSStefano Zampini for (i=0;i<pcis->n;i++) { 42974d379d7bSStefano Zampini PetscInt temp = cxadj[i]; 42984d379d7bSStefano Zampini cxadj[i] = sum; 42994d379d7bSStefano Zampini sum += temp; 43004d379d7bSStefano Zampini } 4301e496cd5dSStefano Zampini cxadj[pcis->n] = sum; 43024d379d7bSStefano Zampini for (i=0;i<graph->ncc;i++) { 43034d379d7bSStefano Zampini for (j=graph->cptr[i];j<graph->cptr[i+1];j++) { 43044d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[j])) { 43054d379d7bSStefano Zampini PetscInt k,sizecc = 0; 43064d379d7bSStefano Zampini for (k=graph->cptr[i];k<graph->cptr[i+1];k++) { 43074d379d7bSStefano Zampini if (PetscBTLookup(is_on_boundary,graph->queue[k])) { 43084d379d7bSStefano Zampini cadjncy[cxadj[graph->queue[j]]+sizecc]=graph->queue[k]; 43094d379d7bSStefano Zampini sizecc++; 43104d379d7bSStefano Zampini } 43114d379d7bSStefano Zampini } 43124d379d7bSStefano Zampini } 43134d379d7bSStefano Zampini } 43144d379d7bSStefano Zampini } 43159b28b941SStefano Zampini if (sum) { 4316e496cd5dSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,pcis->n,cxadj,cadjncy,PETSC_OWN_POINTER);CHKERRQ(ierr); 43174d379d7bSStefano Zampini } else { 43184d379d7bSStefano Zampini ierr = PetscFree(cxadj);CHKERRQ(ierr); 43194d379d7bSStefano Zampini ierr = PetscFree(cadjncy);CHKERRQ(ierr); 43204d379d7bSStefano Zampini } 43214d379d7bSStefano Zampini graph->xadj = 0; 43224d379d7bSStefano Zampini graph->adjncy = 0; 43234d379d7bSStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 43244d379d7bSStefano Zampini ierr = PetscBTDestroy(&is_on_boundary);CHKERRQ(ierr); 43254d379d7bSStefano Zampini } 4326674ae819SStefano Zampini } 43279b28b941SStefano Zampini if (pcbddc->dbg_flag) { 43289b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4329674ae819SStefano Zampini } 4330674ae819SStefano Zampini 433163602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 4332674ae819SStefano Zampini vertex_size = 1; 433363602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 433463602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 433595ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 433663602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4337e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 433863602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4339674ae819SStefano Zampini } 434063602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 434163602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 434263602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4343674ae819SStefano Zampini } 434463602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 4345674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 434663602bcaSStefano Zampini } else { 434763602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 434863602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4349854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 435063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 435163602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 435263602bcaSStefano Zampini } 435363602bcaSStefano Zampini } 4354674ae819SStefano Zampini } 4355674ae819SStefano Zampini 4356674ae819SStefano Zampini /* Setup of Graph */ 4357785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4358e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4359785d1243SStefano Zampini } 4360785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4361e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4362785d1243SStefano Zampini } 436330368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 436430368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 436530368db7SStefano Zampini } 436630368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4367674ae819SStefano Zampini 43684f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 43694f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 43704f1b2e48SStefano Zampini PetscInt *local_subs; 43714f1b2e48SStefano Zampini 43724f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 43734f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 43744f1b2e48SStefano Zampini const PetscInt *idxs; 43754f1b2e48SStefano Zampini PetscInt nl,j; 43764f1b2e48SStefano Zampini 43774f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 43784f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43794f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 43804f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 43814f1b2e48SStefano Zampini } 43824f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43834f1b2e48SStefano Zampini } 43844f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 43854f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 43864f1b2e48SStefano Zampini } 43874f1b2e48SStefano Zampini 4388674ae819SStefano Zampini /* Graph's connected components analysis */ 4389674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4390674ae819SStefano Zampini 4391674ae819SStefano Zampini /* print some info to stdout */ 4392674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4393302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4394674ae819SStefano Zampini } 4395fb180af4SStefano Zampini 4396fb180af4SStefano Zampini /* mark topography has done */ 4397fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 4398674ae819SStefano Zampini PetscFunctionReturn(0); 4399674ae819SStefano Zampini } 4400674ae819SStefano Zampini 4401dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4402674ae819SStefano Zampini #undef __FUNCT__ 4403674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4404dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4405674ae819SStefano Zampini { 4406dc456d91SStefano Zampini PetscSF sf; 4407dc456d91SStefano Zampini PetscLayout map; 4408dc456d91SStefano Zampini const PetscInt *idxs; 4409dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4410dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4411dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4412dc456d91SStefano Zampini PetscMPIInt commsize; 4413674ae819SStefano Zampini PetscBool first_found; 4414674ae819SStefano Zampini PetscErrorCode ierr; 4415674ae819SStefano Zampini 4416674ae819SStefano Zampini PetscFunctionBegin; 4417dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4418dc456d91SStefano Zampini if (subset_mult) { 4419dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4420dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4421dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4422674ae819SStefano Zampini } 4423dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4424dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4425dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4426dc456d91SStefano Zampini for (i=0;i<n;i++) { 4427dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4428dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4429674ae819SStefano Zampini } 4430dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4431dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4432dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4433dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4434dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4435dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4436dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4437dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4438dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4439dc456d91SStefano Zampini 4440dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4441dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4442dc456d91SStefano Zampini if (subset_mult) { 4443dc456d91SStefano Zampini const PetscInt* idxs_mult; 4444dc456d91SStefano Zampini 4445dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4446dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4447dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4448674ae819SStefano Zampini } else { 4449dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4450674ae819SStefano Zampini } 4451dc456d91SStefano Zampini /* local size of new subset */ 4452dc456d91SStefano Zampini n_n = 0; 4453dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4454dc456d91SStefano Zampini 4455dc456d91SStefano Zampini /* global indexes in layout */ 4456dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4457dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4458dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4459dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4460dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4461dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4462dc456d91SStefano Zampini 4463dc456d91SStefano Zampini /* reduce from leaves to roots */ 4464dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 446564a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 446664a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4467dc456d91SStefano Zampini 4468dc456d91SStefano Zampini /* count indexes in local part of layout */ 4469674ae819SStefano Zampini nlocals = 0; 4470674ae819SStefano Zampini first_index = -1; 4471674ae819SStefano Zampini first_found = PETSC_FALSE; 4472dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4473dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4474674ae819SStefano Zampini first_found = PETSC_TRUE; 4475674ae819SStefano Zampini first_index = i; 4476674ae819SStefano Zampini } 4477dc456d91SStefano Zampini nlocals += root_data[i]; 4478674ae819SStefano Zampini } 4479dc456d91SStefano Zampini 4480dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 44815fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4482dc456d91SStefano Zampini start = 0; 448364a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44845fa240b1SStefano Zampini #else 448564a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44865fa240b1SStefano Zampini start = start-nlocals; 44875fa240b1SStefano Zampini #endif 44885fa240b1SStefano Zampini 4489dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4490dc456d91SStefano Zampini *N_n = start + nlocals; 4491dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4492dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4493674ae819SStefano Zampini } 44945fa240b1SStefano Zampini 44955fa240b1SStefano Zampini /* adapt root data with cumulative */ 4496674ae819SStefano Zampini if (first_found) { 4497dc456d91SStefano Zampini PetscInt old_index; 4498dc456d91SStefano Zampini 4499dc456d91SStefano Zampini root_data[first_index] += start; 4500674ae819SStefano Zampini old_index = first_index; 4501dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4502dc456d91SStefano Zampini if (root_data[i]) { 4503dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4504674ae819SStefano Zampini old_index = i; 4505674ae819SStefano Zampini } 4506674ae819SStefano Zampini } 4507674ae819SStefano Zampini } 4508dc456d91SStefano Zampini 4509dc456d91SStefano Zampini /* from roots to leaves */ 4510dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4511dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4512dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4513dc456d91SStefano Zampini 4514dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4515dc456d91SStefano Zampini if (subset_mult) { 4516dc456d91SStefano Zampini const PetscInt* idxs_mult; 4517dc456d91SStefano Zampini PetscInt cum; 4518dc456d91SStefano Zampini 4519dc456d91SStefano Zampini cum = 0; 4520dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4521dc456d91SStefano Zampini for (i=0;i<n;i++) { 4522dc456d91SStefano Zampini PetscInt j; 4523dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4524674ae819SStefano Zampini } 4525dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4526674ae819SStefano Zampini } else { 4527dc456d91SStefano Zampini for (i=0;i<n;i++) { 4528dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4529674ae819SStefano Zampini } 4530674ae819SStefano Zampini } 4531dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4532dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4533674ae819SStefano Zampini PetscFunctionReturn(0); 4534674ae819SStefano Zampini } 45359a7d3425SStefano Zampini 45369a7d3425SStefano Zampini #undef __FUNCT__ 45379a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 45389a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 45399a7d3425SStefano Zampini { 45409a7d3425SStefano Zampini PetscInt i,j; 45419a7d3425SStefano Zampini PetscScalar *alphas; 45429a7d3425SStefano Zampini PetscErrorCode ierr; 45439a7d3425SStefano Zampini 45449a7d3425SStefano Zampini PetscFunctionBegin; 45459a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4546785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 45479a7d3425SStefano Zampini for (i=0;i<n;i++) { 45489a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 45499a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 45509a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 45519a7d3425SStefano Zampini } 45529a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 45539a7d3425SStefano Zampini PetscFunctionReturn(0); 45549a7d3425SStefano Zampini } 45559a7d3425SStefano Zampini 4556e7931f94SStefano Zampini #undef __FUNCT__ 455770cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4558b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4559e7931f94SStefano Zampini { 456052e5ac9dSStefano Zampini IS ranks_send_to; 4561e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4562e7931f94SStefano Zampini PetscMPIInt size,rank,color; 456352e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 456452e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 45653837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 45662b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4567e7931f94SStefano Zampini PetscSubcomm subcomm; 456852e5ac9dSStefano Zampini PetscErrorCode ierr; 4569a57a6d2fSStefano Zampini 4570e7931f94SStefano Zampini PetscFunctionBegin; 45712b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 45722b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 45732b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4574e7931f94SStefano Zampini 4575e7931f94SStefano Zampini /* Get info on mapping */ 45763bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 45773bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4578e7931f94SStefano Zampini 4579e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4580785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4581e7931f94SStefano Zampini xadj[0] = 0; 4582e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4583785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4584785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4585e7931f94SStefano Zampini 45862b510759SStefano Zampini if (threshold) { 4587d023bfaeSStefano Zampini PetscInt xadj_count = 0; 45882b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4589d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4590d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4591d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4592d023bfaeSStefano Zampini xadj_count++; 4593e7931f94SStefano Zampini } 4594e7931f94SStefano Zampini } 4595d023bfaeSStefano Zampini xadj[1] = xadj_count; 4596c8587f34SStefano Zampini } else { 4597e7931f94SStefano Zampini if (xadj[1]) { 4598e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4599e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4600c8587f34SStefano Zampini } 4601e7931f94SStefano Zampini } 46023bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4603e7931f94SStefano Zampini if (use_square) { 4604e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4605e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4606e7931f94SStefano Zampini } 4607e7931f94SStefano Zampini } 4608e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4609e7931f94SStefano Zampini 46103837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4611e7931f94SStefano Zampini 4612e7931f94SStefano Zampini /* 4613e7931f94SStefano Zampini Restrict work on active processes only. 4614e7931f94SStefano Zampini */ 4615e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4616e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4617e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 46182b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4619d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4620e7931f94SStefano Zampini if (color) { 4621e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4622e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4623e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4624c8587f34SStefano Zampini } else { 462552e5ac9dSStefano Zampini Mat subdomain_adj; 462652e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 462752e5ac9dSStefano Zampini MatPartitioning partitioner; 462852e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 462952e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4630b0c7d250SStefano Zampini PetscBool aggregate; 4631b0c7d250SStefano Zampini 4632306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4633785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4634e7931f94SStefano Zampini prank = rank; 4635306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 46368002ef2cSStefano Zampini /* 4637e7931f94SStefano Zampini for (i=0;i<size;i++) { 4638e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4639c8587f34SStefano Zampini } 46408002ef2cSStefano Zampini */ 4641e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4642e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4643c8587f34SStefano Zampini } 4644e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4645b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4646b0c7d250SStefano Zampini if (aggregate) { 4647b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4648b0c7d250SStefano Zampini PetscMPIInt nrank; 4649b0c7d250SStefano Zampini PetscScalar *vals; 4650b0c7d250SStefano Zampini 4651b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4652b0c7d250SStefano Zampini lrows = 0; 4653b0c7d250SStefano Zampini if (nrank<redprocs) { 4654b0c7d250SStefano Zampini lrows = size/redprocs; 4655b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4656b0c7d250SStefano Zampini } 46575fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4658b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4659b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4660b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4661b0c7d250SStefano Zampini row = nrank; 4662b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4663b0c7d250SStefano Zampini cols = adjncy; 4664b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4665b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4666b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4667b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4668b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 466952e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 467052e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 467152e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4672b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4673b0c7d250SStefano Zampini } else { 4674306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4675b0c7d250SStefano Zampini } 467622b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4677e7931f94SStefano Zampini 4678e7931f94SStefano Zampini /* Partition */ 4679306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4680e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4681e7931f94SStefano Zampini if (use_vwgt) { 46823837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4683e7931f94SStefano Zampini v_wgt[0] = local_size; 4684e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4685c8587f34SStefano Zampini } 468628143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 468728143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4688e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4689e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 469022b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4691e7931f94SStefano Zampini 469252e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 469352e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 469452e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 469552e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4696b0c7d250SStefano Zampini if (!redprocs) { 4697b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 469828143c3dSStefano Zampini } else { 4699b0c7d250SStefano Zampini PetscInt idxs[1]; 4700b0c7d250SStefano Zampini PetscMPIInt tag; 4701b0c7d250SStefano Zampini MPI_Request *reqs; 4702b0c7d250SStefano Zampini 4703b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4704b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4705b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4706b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 470728143c3dSStefano Zampini } 4708b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4709b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4710b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4711b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4712e7931f94SStefano Zampini } 471352e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4714e7931f94SStefano Zampini /* clean up */ 4715e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 471652e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4717e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4718e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4719e7931f94SStefano Zampini } 4720e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4721e7931f94SStefano Zampini 4722e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4723e7931f94SStefano Zampini i = 1; 4724e7931f94SStefano Zampini if (color) i=0; 4725e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4726e7931f94SStefano Zampini /* get back IS */ 4727e7931f94SStefano Zampini *is_sends = ranks_send_to; 4728e7931f94SStefano Zampini PetscFunctionReturn(0); 4729e7931f94SStefano Zampini } 4730e7931f94SStefano Zampini 4731e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4732e7931f94SStefano Zampini 4733e7931f94SStefano Zampini #undef __FUNCT__ 4734e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 473553a05cb3SStefano 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[]) 4736e7931f94SStefano Zampini { 473770cf5478SStefano Zampini Mat local_mat; 4738e7931f94SStefano Zampini IS is_sends_internal; 47399d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 474028143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 47419d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4742e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4743e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4744e7931f94SStefano Zampini const PetscInt* is_indices; 4745e7931f94SStefano Zampini MatType new_local_type; 4746e7931f94SStefano Zampini /* buffers */ 4747e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 474828143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 47499d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4750e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4751e7931f94SStefano Zampini /* MPI */ 475228143c3dSStefano Zampini MPI_Comm comm,comm_n; 475328143c3dSStefano Zampini PetscSubcomm subcomm; 4754e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 475528143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 475628143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 475728143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 475828143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 475928143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4760e7931f94SStefano Zampini PetscErrorCode ierr; 4761e7931f94SStefano Zampini 4762e7931f94SStefano Zampini PetscFunctionBegin; 476328143c3dSStefano Zampini /* TODO: add missing checks */ 476428143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 476528143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 476628143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 476728143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4768e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 476928143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4770e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4771e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4772e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4773e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4774e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 477528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 477670cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 477770cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 477828143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 477970cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 478070cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 478170cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 478270cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 478370cf5478SStefano Zampini } 4784e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4785e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4786e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4787e7931f94SStefano Zampini if (!is_sends) { 478828143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4789b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4790c8587f34SStefano Zampini } else { 4791e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4792e7931f94SStefano Zampini is_sends_internal = is_sends; 4793c8587f34SStefano Zampini } 4794e7931f94SStefano Zampini 4795e7931f94SStefano Zampini /* get comm */ 4796a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4797e7931f94SStefano Zampini 4798e7931f94SStefano Zampini /* compute number of sends */ 4799e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4800e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4801e7931f94SStefano Zampini 4802e7931f94SStefano Zampini /* compute number of receives */ 4803e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4804785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4805e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4806e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4807e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4808e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4809e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4810e7931f94SStefano Zampini 481128143c3dSStefano Zampini /* restrict comm if requested */ 481228143c3dSStefano Zampini subcomm = 0; 481328143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 481428143c3dSStefano Zampini if (restrict_comm) { 4815779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4816779c1cceSStefano Zampini 481728143c3dSStefano Zampini color = 0; 481853a05cb3SStefano Zampini if (restrict_full) { 481953a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 482053a05cb3SStefano Zampini } else { 482153a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 482253a05cb3SStefano Zampini } 482328143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 482428143c3dSStefano Zampini subcommsize = commsize - subcommsize; 482528143c3dSStefano Zampini /* check if reuse has been requested */ 482628143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 482728143c3dSStefano Zampini if (*mat_n) { 482828143c3dSStefano Zampini PetscMPIInt subcommsize2; 482928143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 483028143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 483128143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 483228143c3dSStefano Zampini } else { 483328143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 483428143c3dSStefano Zampini } 483528143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4836779c1cceSStefano Zampini PetscMPIInt rank; 4837779c1cceSStefano Zampini 4838779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 483928143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 484028143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 484128143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4842306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 484328143c3dSStefano Zampini } 484428143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 484528143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 484628143c3dSStefano Zampini } else { 484728143c3dSStefano Zampini comm_n = comm; 484828143c3dSStefano Zampini } 484928143c3dSStefano Zampini 4850e7931f94SStefano Zampini /* prepare send/receive buffers */ 4851785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4852e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4853785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4854e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 485528143c3dSStefano Zampini if (nis) { 4856854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 485728143c3dSStefano Zampini } 4858e7931f94SStefano Zampini 485928143c3dSStefano Zampini /* Get data from local matrices */ 4860e7931f94SStefano Zampini if (!isdense) { 4861a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4862e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4863e7931f94SStefano Zampini /* 4864e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4865e7931f94SStefano Zampini send_buffer_idxs should contain: 4866e7931f94SStefano Zampini - MatType_PRIVATE type 4867e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4868e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4869e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4870e7931f94SStefano Zampini */ 4871e7931f94SStefano Zampini } else { 4872e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 48733bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4874854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4875e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4876e7931f94SStefano Zampini send_buffer_idxs[1] = i; 48773bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4878e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 48793bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4880e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4881e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4882e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4883e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4884c8587f34SStefano Zampini } 4885c8587f34SStefano Zampini } 4886e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 488728143c3dSStefano Zampini /* additional is (if any) */ 488828143c3dSStefano Zampini if (nis) { 488928143c3dSStefano Zampini PetscMPIInt psum; 489028143c3dSStefano Zampini PetscInt j; 489128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 489228143c3dSStefano Zampini PetscInt plen; 489328143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 489428143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 489528143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 489628143c3dSStefano Zampini } 4897854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 489828143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 489928143c3dSStefano Zampini PetscInt plen; 490028143c3dSStefano Zampini const PetscInt *is_array_idxs; 490128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 490228143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 490328143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 490428143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 490528143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 490628143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 490728143c3dSStefano Zampini } 490828143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 490928143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 491028143c3dSStefano Zampini } 491128143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 491228143c3dSStefano Zampini } 491328143c3dSStefano Zampini 4914e7931f94SStefano Zampini buf_size_idxs = 0; 4915e7931f94SStefano Zampini buf_size_vals = 0; 491628143c3dSStefano Zampini buf_size_idxs_is = 0; 4917e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4918e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4919e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 492028143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4921e7931f94SStefano Zampini } 4922785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4923785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 492495ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4925e7931f94SStefano Zampini 4926e7931f94SStefano Zampini /* get new tags for clean communications */ 4927e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4928e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 492928143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4930e7931f94SStefano Zampini 4931e7931f94SStefano Zampini /* allocate for requests */ 4932785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4933785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 493495ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4935785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4936785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 493795ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4938e7931f94SStefano Zampini 4939e7931f94SStefano Zampini /* communications */ 4940e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4941e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 494228143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4943e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4944e7931f94SStefano Zampini source_dest = onodes[i]; 4945e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4946e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4947e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4948e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 494928143c3dSStefano Zampini if (nis) { 495028143c3dSStefano 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); 495128143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 495228143c3dSStefano Zampini } 4953e7931f94SStefano Zampini } 4954e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4955e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4956e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4957e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 495828143c3dSStefano Zampini if (nis) { 495928143c3dSStefano 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); 496028143c3dSStefano Zampini } 4961e7931f94SStefano Zampini } 4962e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4963e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4964e7931f94SStefano Zampini 4965e7931f94SStefano Zampini /* assemble new l2g map */ 4966e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4967e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49689d30be91SStefano Zampini new_local_rows = 0; 4969e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49709d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4971e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4972e7931f94SStefano Zampini } 49739d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4974e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49759d30be91SStefano Zampini new_local_rows = 0; 4976e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49779d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 49789d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4979e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4980e7931f94SStefano Zampini } 49819d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 49829d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4983e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4984e7931f94SStefano Zampini 4985e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4986e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4987e7931f94SStefano 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) */ 4988e7931f94SStefano Zampini if (n_recvs) { 498928143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4990e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4991e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4992e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4993e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4994e7931f94SStefano Zampini break; 4995e7931f94SStefano Zampini } 4996e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4997e7931f94SStefano Zampini } 4998e7931f94SStefano Zampini switch (new_local_type_private) { 499928143c3dSStefano Zampini case MATDENSE_PRIVATE: 500028143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5001e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5002e7931f94SStefano Zampini bs = 1; 500328143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 500428143c3dSStefano Zampini new_local_type = MATSEQDENSE; 500528143c3dSStefano Zampini bs = 1; 500628143c3dSStefano Zampini } 5007e7931f94SStefano Zampini break; 5008e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5009e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5010e7931f94SStefano Zampini bs = 1; 5011e7931f94SStefano Zampini break; 5012e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5013e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5014e7931f94SStefano Zampini break; 5015e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5016e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5017e7931f94SStefano Zampini break; 5018e7931f94SStefano Zampini default: 50199d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5020e7931f94SStefano Zampini break; 5021e7931f94SStefano Zampini } 502228143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 502328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 502428143c3dSStefano Zampini bs = 1; 5025e7931f94SStefano Zampini } 5026e7931f94SStefano Zampini 502770cf5478SStefano Zampini /* create MATIS object if needed */ 502870cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5029e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5030e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 503170cf5478SStefano Zampini } else { 503270cf5478SStefano Zampini /* it also destroys the local matrices */ 503370cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 503470cf5478SStefano Zampini } 503570cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5036e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 50379d30be91SStefano Zampini 50389d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 50399d30be91SStefano Zampini 50409d30be91SStefano Zampini /* Global to local map of received indices */ 50419d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 50429d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 50439d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 50449d30be91SStefano Zampini 50459d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 50469d30be91SStefano Zampini buf_size_idxs = 0; 50479d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50489d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 50499d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 50509d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 50519d30be91SStefano Zampini } 50529d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 50539d30be91SStefano Zampini 50549d30be91SStefano Zampini /* set preallocation */ 50559d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 50569d30be91SStefano Zampini if (!newisdense) { 50579d30be91SStefano Zampini PetscInt *new_local_nnz=0; 50589d30be91SStefano Zampini 50599d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 50609d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 50619d30be91SStefano Zampini if (n_recvs) { 50629d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 50639d30be91SStefano Zampini } 50649d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50659d30be91SStefano Zampini PetscInt j; 50669d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 50679d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 50689d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 50699d30be91SStefano Zampini } 50709d30be91SStefano Zampini } else { 50719d30be91SStefano Zampini /* TODO */ 50729d30be91SStefano Zampini } 50739d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 50749d30be91SStefano Zampini } 50759d30be91SStefano Zampini if (new_local_nnz) { 50769d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 50779d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 50789d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 50799d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50809d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 50819d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50829d30be91SStefano Zampini } else { 50839d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50849d30be91SStefano Zampini } 50859d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 50869d30be91SStefano Zampini } else { 50879d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50889d30be91SStefano Zampini } 5089e7931f94SStefano Zampini 5090e7931f94SStefano Zampini /* set values */ 5091e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 50929d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5093e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5094e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5095e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 50969d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5097e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5098e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5099e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 510028143c3dSStefano Zampini } else { 510128143c3dSStefano Zampini /* TODO */ 5102e7931f94SStefano Zampini } 5103e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5104e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5105e7931f94SStefano Zampini } 5106e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5107e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 510870cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 510970cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 51109d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 51119d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5112e7931f94SStefano Zampini 5113dfd14d43SStefano Zampini #if 0 511428143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5115e7931f94SStefano Zampini Vec lvec,rvec; 5116e7931f94SStefano Zampini PetscReal infty_error; 5117e7931f94SStefano Zampini 51182a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5119e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5120e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5121e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 512270cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5123e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5124e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5125e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5126e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5127e7931f94SStefano Zampini } 512828143c3dSStefano Zampini #endif 5129e7931f94SStefano Zampini 513028143c3dSStefano Zampini /* assemble new additional is (if any) */ 513128143c3dSStefano Zampini if (nis) { 513228143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 513328143c3dSStefano Zampini 513428143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5135854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 513628143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 513728143c3dSStefano Zampini psum = 0; 513828143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 513928143c3dSStefano Zampini for (j=0;j<nis;j++) { 514028143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 514128143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 514228143c3dSStefano Zampini psum += plen; 514328143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 514428143c3dSStefano Zampini } 514528143c3dSStefano Zampini } 5146854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5147854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 514828143c3dSStefano Zampini for (i=1;i<nis;i++) { 514928143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 515028143c3dSStefano Zampini } 515128143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 515228143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 515328143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 515428143c3dSStefano Zampini for (j=0;j<nis;j++) { 515528143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 515628143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 515728143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 515828143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 515928143c3dSStefano Zampini } 516028143c3dSStefano Zampini } 516128143c3dSStefano Zampini for (i=0;i<nis;i++) { 516228143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 516328143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 516428143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 516528143c3dSStefano Zampini } 516628143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 516728143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 516828143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 516928143c3dSStefano Zampini } 5170e7931f94SStefano Zampini /* free workspace */ 517128143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5172e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5173e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5174e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5175e7931f94SStefano Zampini if (isdense) { 5176e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5177e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5178e7931f94SStefano Zampini } else { 5179e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5180e7931f94SStefano Zampini } 518128143c3dSStefano Zampini if (nis) { 518228143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 518328143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 518428143c3dSStefano Zampini } 5185e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5186e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 518728143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5188e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5189e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 519028143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5191e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5192e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5193e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5194e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5195e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 519628143c3dSStefano Zampini if (nis) { 519728143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 519828143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 519928143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 520028143c3dSStefano Zampini } 520128143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 520228143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 520328143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 520428143c3dSStefano Zampini for (i=0;i<nis;i++) { 520528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 520628143c3dSStefano Zampini } 520753a05cb3SStefano Zampini *mat_n = NULL; 520828143c3dSStefano Zampini } 5209e7931f94SStefano Zampini PetscFunctionReturn(0); 5210e7931f94SStefano Zampini } 5211a57a6d2fSStefano Zampini 521212edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5213af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 521412edc857SStefano Zampini 5215c8587f34SStefano Zampini #undef __FUNCT__ 5216c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5217c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5218c8587f34SStefano Zampini { 5219c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5220c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 522120a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 52229881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 522320a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 52246e683305SStefano Zampini IS coarse_is,*isarray; 52256e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 522630368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5227f9eb5b7dSStefano Zampini PC pc_temp; 5228c8587f34SStefano Zampini PCType coarse_pc_type; 5229c8587f34SStefano Zampini KSPType coarse_ksp_type; 5230f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 52314f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 52326e683305SStefano Zampini Mat t_coarse_mat_is; 52336e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 52346e683305SStefano Zampini PetscMPIInt all_procs; 523574e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 523668457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 523722bc73bbSStefano Zampini PetscScalar *array; 52389881197aSStefano Zampini PetscErrorCode ierr; 5239fdc09c96SStefano Zampini 5240c8587f34SStefano Zampini PetscFunctionBegin; 5241c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 524268457ee5SStefano 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 */ 5243fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 52445a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5245fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5246f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5247f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5248f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5249fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 525051bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 525151bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5252dc4bcba2SStefano Zampini PC pc; 5253dc4bcba2SStefano Zampini PetscBool isbddc; 5254dc4bcba2SStefano Zampini 5255dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5256dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5257dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5258dc4bcba2SStefano Zampini if (isbddc) { 5259dc4bcba2SStefano Zampini ierr = PCDestroy(&pc);CHKERRQ(ierr); 5260dc4bcba2SStefano Zampini } 5261727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 5262fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5263fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5264fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5265f4ddd8eeSStefano Zampini } 5266fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5267fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5268f4ddd8eeSStefano Zampini } 526970cf5478SStefano Zampini /* reset any subassembling information */ 527070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 52716e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 52726e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5273fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5274f4ddd8eeSStefano Zampini } 5275c8587f34SStefano Zampini 52766e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 52772b510759SStefano Zampini im_active = !!(pcis->n); 52782b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 52796e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 52806e683305SStefano Zampini void_procs = all_procs-active_procs; 52816e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 528274e2c79eSStefano Zampini redist = PETSC_FALSE; 528322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 52846e683305SStefano Zampini csin_ml = PETSC_TRUE; 52856e683305SStefano Zampini ncoarse_ml = void_procs; 5286779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5287779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 52886e683305SStefano Zampini csin_ds = PETSC_TRUE; 528918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 529018a45a71SStefano Zampini redist = PETSC_TRUE; 529118a45a71SStefano Zampini } else { 52926e683305SStefano Zampini csin_ds = PETSC_TRUE; 5293779c1cceSStefano Zampini ncoarse_ds = active_procs; 5294779c1cceSStefano Zampini redist = PETSC_TRUE; 529518a45a71SStefano Zampini } 52966e683305SStefano Zampini } else { 52976e683305SStefano Zampini csin_ml = PETSC_FALSE; 52986e683305SStefano Zampini ncoarse_ml = all_procs; 52996e683305SStefano Zampini if (void_procs) { 53006e683305SStefano Zampini csin_ds = PETSC_TRUE; 53016e683305SStefano Zampini ncoarse_ds = void_procs; 53026e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 53036e683305SStefano Zampini } else { 5304779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 530574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 530674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 530774e2c79eSStefano Zampini redist = PETSC_TRUE; 530874e2c79eSStefano Zampini } else { 53096e683305SStefano Zampini csin_ds = PETSC_FALSE; 53106e683305SStefano Zampini ncoarse_ds = all_procs; 53116e683305SStefano Zampini } 53126e683305SStefano Zampini } 531374e2c79eSStefano Zampini } 53146e683305SStefano Zampini 53156e683305SStefano Zampini /* 53166e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 53176e683305SStefano Zampini - we have not exceeded the number of levels requested 53186e683305SStefano Zampini - we can actually subassemble the active processes 53196e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 53206e683305SStefano Zampini */ 53216e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 53226e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 53236e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 53246e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 53256e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5326f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 53272b510759SStefano Zampini } else { 5328f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5329c8587f34SStefano Zampini } 5330c8587f34SStefano Zampini } 53316e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 53326e683305SStefano Zampini if (multilevel_allowed) { 53336e683305SStefano Zampini ncoarse = ncoarse_ml; 53346e683305SStefano Zampini csin = csin_ml; 533558da7f69SStefano Zampini redist = PETSC_FALSE; 53366e683305SStefano Zampini } else { 53376e683305SStefano Zampini ncoarse = ncoarse_ds; 53386e683305SStefano Zampini csin = csin_ds; 53396e683305SStefano Zampini } 5340e7931f94SStefano Zampini 5341abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5342abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5343abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5344abbbba34SStefano Zampini 5345abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 534622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 534722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 534822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 534922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5350e176bc59SStefano 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); 53516e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 53526e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53536e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5354abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5355abbbba34SStefano Zampini 53566e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 535730368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 53586e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 53596e683305SStefano Zampini const PetscInt *idxs; 53606e683305SStefano Zampini ISLocalToGlobalMapping tmap; 53616e683305SStefano Zampini 53626e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 53630be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 53646e683305SStefano Zampini /* allocate space for temporary storage */ 5365854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5366854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 53676e683305SStefano Zampini /* allocate for IS array */ 53686e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 53696e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 537030368db7SStefano Zampini nisvert = 0; 537130368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 537230368db7SStefano Zampini nisvert = 1; 537330368db7SStefano Zampini } 537430368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5375854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 53766e683305SStefano Zampini /* dofs splitting */ 53776e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 53786e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 53796e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 53806e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53816e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53826e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53836e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 538430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 53856e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 53866e683305SStefano Zampini } 53876e683305SStefano Zampini /* neumann boundaries */ 53886e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 53896e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 53906e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 53916e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53926e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53936e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&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[nisdofs]);CHKERRQ(ierr); 53966e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 53976e683305SStefano Zampini } 539830368db7SStefano Zampini /* primal vertices (benign) */ 539930368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 540030368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 540130368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 540230368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 540330368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 540430368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 540530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 540630368db7SStefano Zampini } 54076e683305SStefano Zampini /* free memory */ 54086e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 54096e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 54106e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 54116e683305SStefano Zampini } else { 54126e683305SStefano Zampini nis = 0; 54136e683305SStefano Zampini nisdofs = 0; 54146e683305SStefano Zampini nisneu = 0; 541530368db7SStefano Zampini nisvert = 0; 54166e683305SStefano Zampini isarray = NULL; 54176e683305SStefano Zampini } 54186e683305SStefano Zampini /* destroy no longer needed map */ 54196e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 54206e683305SStefano Zampini 54216e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 54226e683305SStefano Zampini coarse_mat_is = NULL; 54236e683305SStefano Zampini if (csin) { 54246e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 542574e2c79eSStefano Zampini if (redist) { 542674e2c79eSStefano Zampini PetscMPIInt rank; 5427779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 542874e2c79eSStefano Zampini 542974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 543058da7f69SStefano Zampini spc = active_procs/ncoarse; 543158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5432779c1cceSStefano Zampini if (im_active) { 5433779c1cceSStefano Zampini destsize = 1; 543474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 543574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 543674e2c79eSStefano Zampini } else { 543774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 543874e2c79eSStefano Zampini } 543974e2c79eSStefano Zampini } else { 5440779c1cceSStefano Zampini destsize = 0; 54416e683305SStefano Zampini } 5442779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5443779c1cceSStefano Zampini } else if (csin_type_simple) { 54446e683305SStefano Zampini PetscMPIInt rank; 54456e683305SStefano Zampini PetscInt issize,isidx; 5446779c1cceSStefano Zampini 54476e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 54486e683305SStefano Zampini if (im_active) { 54496e683305SStefano Zampini issize = 1; 54506e683305SStefano Zampini isidx = (PetscInt)rank; 54516e683305SStefano Zampini } else { 54526e683305SStefano Zampini issize = 0; 54536e683305SStefano Zampini isidx = -1; 54546e683305SStefano Zampini } 54556e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5456779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5457b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54586e683305SStefano Zampini } 5459779c1cceSStefano Zampini 5460779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5461779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5462779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5463779c1cceSStefano Zampini PetscInt *coarse_candidates; 5464779c1cceSStefano Zampini const PetscInt* tisindices; 5465779c1cceSStefano Zampini 5466779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5467779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5468779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5469779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5470779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5471779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5472779c1cceSStefano Zampini } 5473779c1cceSStefano Zampini } 5474779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5475779c1cceSStefano Zampini 5476779c1cceSStefano Zampini 54776e683305SStefano Zampini if (pcbddc->dbg_flag) { 54786e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54796e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 54806e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54816e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5482779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 54836e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 54846e683305SStefano Zampini } 54856e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 54866e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54876e683305SStefano Zampini } 54886e683305SStefano Zampini /* shift the pattern on coarse candidates */ 54896e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 54906e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5491854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 54926e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 54936e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 54946e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 54956e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 54966e683305SStefano Zampini } 54976e683305SStefano Zampini if (pcbddc->dbg_flag) { 54986e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54996e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 55006e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55016e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55026e683305SStefano Zampini } 5503779c1cceSStefano Zampini } 55046e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 550553a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 550653a05cb3SStefano 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); 550753a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 550853a05cb3SStefano 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); 550953a05cb3SStefano Zampini } 55106e683305SStefano Zampini } else { 55116e683305SStefano Zampini if (pcbddc->dbg_flag) { 55126e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55136e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 55146e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55156e683305SStefano Zampini } 55166e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 55176e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 55186e683305SStefano Zampini } 55196e683305SStefano Zampini 55206e683305SStefano Zampini /* create local to global scatters for coarse problem */ 552168457ee5SStefano Zampini if (compute_vecs) { 55226e683305SStefano Zampini PetscInt lrows; 55236e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 55246e683305SStefano Zampini if (coarse_mat_is) { 55256e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 55266e683305SStefano Zampini } else { 55276e683305SStefano Zampini lrows = 0; 55286e683305SStefano Zampini } 55296e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 55306e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 55316e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 55326e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55336e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55346e683305SStefano Zampini } 55356e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 55366e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5537c8587f34SStefano Zampini 5538f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5539f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5540f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5541f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5542f9eb5b7dSStefano Zampini } else { 5543f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5544f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5545c8587f34SStefano Zampini } 5546c8587f34SStefano Zampini 55476e683305SStefano Zampini /* print some info if requested */ 55486e683305SStefano Zampini if (pcbddc->dbg_flag) { 55496e683305SStefano Zampini if (!multilevel_allowed) { 55506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55516e683305SStefano Zampini if (multilevel_requested) { 55526e683305SStefano 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); 55536e683305SStefano Zampini } else if (pcbddc->max_levels) { 55546e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 55556e683305SStefano Zampini } 55566e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55576e683305SStefano Zampini } 55586e683305SStefano Zampini } 55596e683305SStefano Zampini 5560f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 55616e683305SStefano Zampini if (coarse_mat_is) { 55626e683305SStefano Zampini MatReuse coarse_mat_reuse; 55636a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 55646e683305SStefano Zampini if (pcbddc->dbg_flag) { 55656e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 55666e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 55676e683305SStefano Zampini } 5568f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5569312be037SStefano Zampini char prefix[256],str_level[16]; 5570e604994aSStefano Zampini size_t len; 55716e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5572422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5573c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5574f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 55755f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5576c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 55776e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5578c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5579c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5580e604994aSStefano Zampini /* prefix */ 5581e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5582e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5583e604994aSStefano Zampini if (!pcbddc->current_level) { 5584e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5585e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5586c8587f34SStefano Zampini } else { 5587e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5588312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5589312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 559034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5591312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5592e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5593e604994aSStefano Zampini } 5594e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 55953e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 55963e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 55973e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 55983e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5599f9eb5b7dSStefano Zampini /* allow user customization */ 5600f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 56013e3c6dadSStefano Zampini } 56023e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 560351bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 56043e3c6dadSStefano Zampini if (nisdofs) { 56053e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 56063e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 56073e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 56083e3c6dadSStefano Zampini } 56093e3c6dadSStefano Zampini } 56103e3c6dadSStefano Zampini if (nisneu) { 56113e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 56123e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5613312be037SStefano Zampini } 561430368db7SStefano Zampini if (nisvert) { 561530368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 561630368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 561730368db7SStefano Zampini } 5618f9eb5b7dSStefano Zampini 5619f9eb5b7dSStefano Zampini /* get some info after set from options */ 5620f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5621f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 56224f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 56236e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5624f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5625f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5626f9eb5b7dSStefano Zampini } 562739f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 56284f3a063dSStefano Zampini if (isredundant) { 56294f3a063dSStefano Zampini KSP inner_ksp; 56304f3a063dSStefano Zampini PC inner_pc; 56314f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 56324f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 56334f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 56344f3a063dSStefano Zampini } 5635f9eb5b7dSStefano Zampini 5636f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5637fa7f1dd8SStefano Zampini if (coarse_reuse) { 563881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5639fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 56406e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5641fa7f1dd8SStefano Zampini } else { 56426e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5643fa7f1dd8SStefano Zampini } 5644c8587f34SStefano Zampini if (isbddc || isnn) { 564522bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 564670cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5647b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 564822b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 56496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56506e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 56516e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 56526e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 565322b6e8a2SStefano Zampini } 565470cf5478SStefano Zampini } 565553a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 565670cf5478SStefano Zampini } else { 565722bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 565822bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 565922bc73bbSStefano Zampini } 566022bc73bbSStefano Zampini } else { 56612e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5662c8587f34SStefano Zampini } 5663c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5664c8587f34SStefano Zampini 56653301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 56665a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 56673301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 56683301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 56693301b35fSStefano Zampini } 56703301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 56713301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 56723301b35fSStefano Zampini } 56733301b35fSStefano Zampini if (pc->pmat->spd_set) { 56743301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 56753301b35fSStefano Zampini } 56766e683305SStefano Zampini /* set operators */ 56775f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 56786e683305SStefano Zampini if (pcbddc->dbg_flag) { 56796e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56806e683305SStefano Zampini } 56816e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 56826e683305SStefano Zampini coarse_mat = 0; 56836e683305SStefano Zampini } 56846e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5685b1ecc7b1SStefano Zampini #if 0 5686b9b85e73SStefano Zampini { 5687b9b85e73SStefano Zampini PetscViewer viewer; 5688b9b85e73SStefano Zampini char filename[256]; 5689b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5690b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5691b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5692b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5693b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5694b9b85e73SStefano Zampini } 5695b9b85e73SStefano Zampini #endif 5696c8587f34SStefano Zampini 5697c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5698298c0119SStefano Zampini #if 0 5699c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5700c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 570198a51de6SStefano Zampini } 5702298c0119SStefano Zampini #endif 5703b0f5fe93SStefano Zampini /* hack */ 570498a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 570598a51de6SStefano Zampini Vec crhs,csol; 570604708bb6SStefano Zampini 5707f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5708f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5709f347579bSStefano Zampini if (!csol) { 57102a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5711f9eb5b7dSStefano Zampini } 5712f347579bSStefano Zampini if (!crhs) { 57132a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5714f347579bSStefano Zampini } 5715b0f5fe93SStefano Zampini } 5716b0f5fe93SStefano Zampini 5717b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5718b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5719b0f5fe93SStefano Zampini 5720b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 57214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 57224f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 57234f1b2e48SStefano Zampini } 5724b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5725b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5726b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5727b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5728b0f5fe93SStefano Zampini if (coarse_mat) { 5729b0f5fe93SStefano Zampini Vec nullv; 5730b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5731b0f5fe93SStefano Zampini PetscInt nl; 5732b0f5fe93SStefano Zampini 5733b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5734b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5735b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5736b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5737b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5738b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5739b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5740b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5741b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5742b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5743b0f5fe93SStefano Zampini } 5744b0f5fe93SStefano Zampini } 5745b0f5fe93SStefano Zampini 5746b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5747b0f5fe93SStefano Zampini PetscBool ispreonly; 5748b0f5fe93SStefano Zampini 5749b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5750b0f5fe93SStefano Zampini PetscBool isnull; 5751b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5752b0f5fe93SStefano Zampini if (isnull) { 575330368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5754b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5755b0f5fe93SStefano Zampini } else { 5756b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5757b0f5fe93SStefano Zampini } 5758b0f5fe93SStefano Zampini } else { 5759b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5760b0f5fe93SStefano Zampini } 5761b0f5fe93SStefano Zampini } 5762b0f5fe93SStefano Zampini /* setup coarse ksp */ 5763b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5764cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5765cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 57666e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5767c8587f34SStefano Zampini KSP check_ksp; 57682b510759SStefano Zampini KSPType check_ksp_type; 5769c8587f34SStefano Zampini PC check_pc; 57706e683305SStefano Zampini Vec check_vec,coarse_vec; 57716a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 57722b510759SStefano Zampini PetscInt its; 57736e683305SStefano Zampini PetscBool compute_eigs; 57746e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 57756e683305SStefano Zampini PetscInt neigs; 57768e185a42SStefano Zampini const char *prefix; 5777c8587f34SStefano Zampini 57782b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 57796e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5780422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 578123ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5782f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 57832b510759SStefano Zampini if (ispreonly) { 57842b510759SStefano Zampini check_ksp_type = KSPPREONLY; 57856e683305SStefano Zampini compute_eigs = PETSC_FALSE; 57862b510759SStefano Zampini } else { 5787cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 57886e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5789c8587f34SStefano Zampini } 5790c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 57916e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 57926e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 57936e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5794a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5795a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5796a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5797a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5798c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5799c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5800c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5801c8587f34SStefano Zampini /* create random vec */ 58026e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 58036e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5804c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5805c8587f34SStefano Zampini if (CoarseNullSpace) { 5806c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5807c8587f34SStefano Zampini } 58086e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5809c8587f34SStefano Zampini /* solve coarse problem */ 58106e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5811c8587f34SStefano Zampini if (CoarseNullSpace) { 58126e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5813c8587f34SStefano Zampini } 5814cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 58156e683305SStefano Zampini if (compute_eigs) { 5816854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5817854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 58186e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 58196e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 58206e683305SStefano Zampini lambda_min = eigs_r[0]; 58216e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 58226e683305SStefano Zampini if (lambda_max>lambda_min) { 5823cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5824cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5825cbcc2c2aSStefano Zampini } 5826c8587f34SStefano Zampini } 5827c8587f34SStefano Zampini } 5828cbcc2c2aSStefano Zampini 5829c8587f34SStefano Zampini /* check coarse problem residual error */ 58306e683305SStefano Zampini if (pcbddc->dbg_flag) { 58316e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 58326e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58336e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5834c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 58356e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 58366e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5837c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5838779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 58396e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 58406e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 58416e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 58426e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5843b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5844b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5845b0f5fe93SStefano Zampini } 58466e683305SStefano Zampini if (compute_eigs) { 58476e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5848deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5849c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 58506e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 58516e683305SStefano 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); 58526e683305SStefano Zampini for (i=0;i<neigs;i++) { 58536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5854c8587f34SStefano Zampini } 58556e683305SStefano Zampini } 58566e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 58576e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58586e683305SStefano Zampini } 5859c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 58606e683305SStefano Zampini if (compute_eigs) { 58616e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 58626e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5863c8587f34SStefano Zampini } 58646e683305SStefano Zampini } 58656e683305SStefano Zampini } 5866cbcc2c2aSStefano Zampini /* print additional info */ 5867cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 58686e683305SStefano Zampini /* waits until all processes reaches this point */ 58696e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5870cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5871cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5872cbcc2c2aSStefano Zampini } 5873cbcc2c2aSStefano Zampini 58742b510759SStefano Zampini /* free memory */ 5875c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5876fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5877c8587f34SStefano Zampini PetscFunctionReturn(0); 5878c8587f34SStefano Zampini } 5879674ae819SStefano Zampini 5880f34684f1SStefano Zampini #undef __FUNCT__ 5881f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5882f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5883f34684f1SStefano Zampini { 5884f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5885f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5886f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5887dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5888dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 588973be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5890dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5891f34684f1SStefano Zampini PetscErrorCode ierr; 5892f34684f1SStefano Zampini 5893f34684f1SStefano Zampini PetscFunctionBegin; 5894f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 58950e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 58960e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5897727cdba6SStefano Zampini } 5898dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 58993bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5900dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5901dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5902dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5903dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5904dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5905dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 59060e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 59070e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 59080e6343abSStefano Zampini } 5909dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5910dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5911dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5912dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5913dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5914f34684f1SStefano Zampini 5915f34684f1SStefano Zampini /* check numbering */ 5916f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5917019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5918dc456d91SStefano Zampini PetscInt i; 5919b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5920f34684f1SStefano Zampini 5921f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5922f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5923f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 59241575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5925019a44ceSStefano Zampini /* counter */ 5926019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5927019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5928019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5929019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5930019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5931019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5932f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5933f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5934727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5935f34684f1SStefano Zampini } 5936f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5937f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5938f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5939e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5940e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5941e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5942e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5943f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5944019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5945f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5946019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 594775c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 594875c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5949b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5950019a44ceSStefano 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); 5951f34684f1SStefano Zampini } 5952f34684f1SStefano Zampini } 5953019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5954b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5955f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5956f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5957f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5958f34684f1SStefano Zampini } 5959f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5960f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5961e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5962e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5963f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5964f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5965b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5966ca8b9ea9SStefano Zampini PetscInt *gidxs; 5967ca8b9ea9SStefano Zampini 5968ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 59693bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5970f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5971f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5972f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5973f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 59744bc2dc4bSStefano 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); 5975f34684f1SStefano Zampini } 5976f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5977ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5978f34684f1SStefano Zampini } 5979f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59801575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5981302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5982f34684f1SStefano Zampini } 59838bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5984f34684f1SStefano Zampini /* get back data */ 5985f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5986f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5987674ae819SStefano Zampini PetscFunctionReturn(0); 5988674ae819SStefano Zampini } 5989674ae819SStefano Zampini 5990e456f2a8SStefano Zampini #undef __FUNCT__ 5991e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5992a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5993e456f2a8SStefano Zampini { 5994e456f2a8SStefano Zampini IS localis_t; 5995a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5996e456f2a8SStefano Zampini PetscScalar *vals; 5997e456f2a8SStefano Zampini PetscErrorCode ierr; 5998e456f2a8SStefano Zampini 5999e456f2a8SStefano Zampini PetscFunctionBegin; 6000a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6001e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6002854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6003e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6004e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6005a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6006a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 60071035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6008a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 60091035eff8SStefano Zampini } 6010a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6011e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6012e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6013a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6014a7dc3881SStefano Zampini /* now compute set in local ordering */ 6015a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6016a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6017a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6018a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6019a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6020ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6021e456f2a8SStefano Zampini lsize++; 6022e456f2a8SStefano Zampini } 6023e456f2a8SStefano Zampini } 6024854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6025a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6026ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6027e456f2a8SStefano Zampini idxs[lsize++] = i; 6028e456f2a8SStefano Zampini } 6029e456f2a8SStefano Zampini } 6030a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6031a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6032e456f2a8SStefano Zampini *localis = localis_t; 6033e456f2a8SStefano Zampini PetscFunctionReturn(0); 6034e456f2a8SStefano Zampini } 6035906d46d4SStefano Zampini 6036906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6037906d46d4SStefano Zampini #undef __FUNCT__ 6038906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6039906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6040906d46d4SStefano Zampini { 6041906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6042906d46d4SStefano Zampini PetscErrorCode ierr; 6043906d46d4SStefano Zampini 6044906d46d4SStefano Zampini PetscFunctionBegin; 6045906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6046906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6047906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6048906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6049906d46d4SStefano Zampini PetscFunctionReturn(0); 6050906d46d4SStefano Zampini } 6051906d46d4SStefano Zampini 6052906d46d4SStefano Zampini #undef __FUNCT__ 6053906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6054906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6055906d46d4SStefano Zampini { 6056906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6057906d46d4SStefano Zampini PetscErrorCode ierr; 6058906d46d4SStefano Zampini 6059906d46d4SStefano Zampini PetscFunctionBegin; 6060906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6061906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6062906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6063906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6064906d46d4SStefano Zampini PetscFunctionReturn(0); 6065906d46d4SStefano Zampini } 6066b96c3477SStefano Zampini 6067b96c3477SStefano Zampini #undef __FUNCT__ 6068b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 606908122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6070b96c3477SStefano Zampini { 6071a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6072b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6073b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6074a64f4aa4SStefano Zampini Mat S_j; 6075b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6076b96c3477SStefano Zampini PetscBool free_used_adj; 6077b96c3477SStefano Zampini PetscErrorCode ierr; 6078b96c3477SStefano Zampini 6079b96c3477SStefano Zampini PetscFunctionBegin; 6080b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6081b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 608208122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6083b96c3477SStefano Zampini used_xadj = NULL; 6084b96c3477SStefano Zampini used_adjncy = NULL; 6085b96c3477SStefano Zampini } else { 608608122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 608708122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 608808122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 608908122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6090b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6091b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6092b96c3477SStefano Zampini } else { 60932fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6094b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6095b96c3477SStefano Zampini PetscInt nvtxs; 6096b96c3477SStefano Zampini 60972fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 60982fffb893SStefano Zampini if (flg_row) { 6099b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6100b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6101b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6102b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 61032fffb893SStefano Zampini } else { 61042fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 61052fffb893SStefano Zampini used_xadj = NULL; 61062fffb893SStefano Zampini used_adjncy = NULL; 61072fffb893SStefano Zampini } 61082fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6109b96c3477SStefano Zampini } 6110b96c3477SStefano Zampini } 6111d5574798SStefano Zampini 6112d5574798SStefano Zampini /* setup sub_schurs data */ 6113a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6114*df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6115*df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6116a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6117ca92afb2SStefano 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); 6118a64f4aa4SStefano Zampini } else { 61196816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 612004708bb6SStefano Zampini PetscBool isseqaij; 6121a3df083aSStefano Zampini PetscInt benign_n; 6122a3df083aSStefano Zampini 61235feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 61245feab87aSStefano Zampini PetscInt n_vertices; 61255feab87aSStefano Zampini 61265feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 61272034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 61285feab87aSStefano Zampini } 612904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 613004708bb6SStefano Zampini if (!isseqaij) { 613104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 613204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 613304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 613404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 613504708bb6SStefano Zampini } else { 613604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 613704708bb6SStefano Zampini } 613804708bb6SStefano Zampini } 6139a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6140a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6141ca92afb2SStefano Zampini } else { 6142a3df083aSStefano Zampini benign_n = 0; 6143ca92afb2SStefano Zampini } 6144a3df083aSStefano 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); 6145ca92afb2SStefano Zampini } 6146b96c3477SStefano Zampini 6147b96c3477SStefano Zampini /* free adjacency */ 6148b96c3477SStefano Zampini if (free_used_adj) { 6149b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6150b96c3477SStefano Zampini } 6151b96c3477SStefano Zampini PetscFunctionReturn(0); 6152b96c3477SStefano Zampini } 6153b96c3477SStefano Zampini 6154b96c3477SStefano Zampini #undef __FUNCT__ 6155b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 615608122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6157b96c3477SStefano Zampini { 6158b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6159b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6160b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6161b96c3477SStefano Zampini PCBDDCGraph graph; 6162b96c3477SStefano Zampini PetscErrorCode ierr; 6163b96c3477SStefano Zampini 6164b96c3477SStefano Zampini PetscFunctionBegin; 6165b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 616608122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 61673301b35fSStefano Zampini IS verticesIS,verticescomm; 61683301b35fSStefano Zampini PetscInt vsize,*idxs; 6169b96c3477SStefano Zampini 6170b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 61713301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 61723301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61733301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 61743301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 61753301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6176b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 61777fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 61783301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 61793301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6180b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6181b96c3477SStefano Zampini /* 6182b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6183b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6184b96c3477SStefano Zampini } 6185b96c3477SStefano Zampini */ 6186b96c3477SStefano Zampini } else { 6187b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6188b96c3477SStefano Zampini } 6189b96c3477SStefano Zampini 6190b96c3477SStefano Zampini /* sub_schurs init */ 6191a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6192a64f4aa4SStefano Zampini 6193b96c3477SStefano Zampini /* free graph struct */ 619408122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6195b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6196b96c3477SStefano Zampini } 6197b96c3477SStefano Zampini PetscFunctionReturn(0); 6198b96c3477SStefano Zampini } 6199fa34dd3eSStefano Zampini 6200fa34dd3eSStefano Zampini #undef __FUNCT__ 6201fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6202fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6203fa34dd3eSStefano Zampini { 6204fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6205fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6206fa34dd3eSStefano Zampini PetscErrorCode ierr; 6207fa34dd3eSStefano Zampini 6208fa34dd3eSStefano Zampini PetscFunctionBegin; 6209fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6210fa34dd3eSStefano Zampini IS zerodiag = NULL; 62114f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6212fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 62134f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 621475c01103SStefano Zampini PetscReal norm; 6215fa34dd3eSStefano Zampini PetscInt i; 6216fa34dd3eSStefano Zampini 6217fa34dd3eSStefano Zampini /* B0 and B0_B */ 6218fa34dd3eSStefano Zampini if (zerodiag) { 6219fa34dd3eSStefano Zampini IS dummy; 6220fa34dd3eSStefano Zampini 62214f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 62224f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6223fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6224fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6225fa34dd3eSStefano Zampini } 6226fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6227fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6228fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6229fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6230fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6231fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6232fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6233fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6234fa34dd3eSStefano Zampini /* S_j */ 6235fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6236fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6237fa34dd3eSStefano Zampini 6238fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6239fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6240fa34dd3eSStefano Zampini /* continuous in primal space */ 6241fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6242fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6243fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6244fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 62454f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 62464f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6247fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6248fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6249fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6250fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6251fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6252fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6253fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6254fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6255fa34dd3eSStefano Zampini 6256fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6257fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6258fa34dd3eSStefano Zampini /* local with Schur */ 6259fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6260fa34dd3eSStefano Zampini if (zerodiag) { 6261fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62624f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6263fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6264fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6265fa34dd3eSStefano Zampini } 6266fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6267fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6268fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6269fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6270fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6271fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6272fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6273fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6274fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6275fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6276fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6277fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6278fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6279fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6280fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6281fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6282fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6283fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6284fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6285fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6286fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6287fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6288fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6289fa34dd3eSStefano Zampini if (zerodiag) { 6290fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6291fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62924f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6293fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6294fa34dd3eSStefano Zampini } 6295fa34dd3eSStefano Zampini /* BDDC */ 6296fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6297fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6298fa34dd3eSStefano Zampini 6299fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6300fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6301fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6302fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 63034f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 63044f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6305fa34dd3eSStefano Zampini } 63064f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6307fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6308fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6309fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6310fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6311fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6312fa34dd3eSStefano Zampini } 6313fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6314fa34dd3eSStefano Zampini } 6315