11cf9b237SStefano Zampini #include <../src/mat/impls/aij/seq/aij.h> 2ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddc.h> 3ab5c6b0cSJed Brown #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 4674ae819SStefano Zampini #include <petscblaslapack.h> 5674ae819SStefano Zampini 6906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y); 7906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y); 8906d46d4SStefano Zampini 9a3df083aSStefano Zampini #undef __FUNCT__ 10a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private_Private" 11a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose) 12a3df083aSStefano Zampini { 13a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 14a3df083aSStefano Zampini PetscErrorCode ierr; 15a3df083aSStefano Zampini PetscBool apply_right,apply_left,reset_x; 16a3df083aSStefano Zampini 17a3df083aSStefano Zampini PetscFunctionBegin; 18a3df083aSStefano Zampini ierr = MatShellGetContext(A,&ctx);CHKERRQ(ierr); 19a3df083aSStefano Zampini if (transpose) { 20a3df083aSStefano Zampini apply_right = ctx->apply_left; 21a3df083aSStefano Zampini apply_left = ctx->apply_right; 22a3df083aSStefano Zampini } else { 23a3df083aSStefano Zampini apply_right = ctx->apply_right; 24a3df083aSStefano Zampini apply_left = ctx->apply_left; 25a3df083aSStefano Zampini } 26a3df083aSStefano Zampini reset_x = PETSC_FALSE; 27a3df083aSStefano Zampini if (apply_right) { 28a3df083aSStefano Zampini const PetscScalar *ax; 29a3df083aSStefano Zampini PetscInt nl,i; 30a3df083aSStefano Zampini 31a3df083aSStefano Zampini ierr = VecGetLocalSize(x,&nl);CHKERRQ(ierr); 32a3df083aSStefano Zampini ierr = VecGetArrayRead(x,&ax);CHKERRQ(ierr); 33a3df083aSStefano Zampini ierr = PetscMemcpy(ctx->work,ax,nl*sizeof(PetscScalar));CHKERRQ(ierr); 34a3df083aSStefano Zampini ierr = VecRestoreArrayRead(x,&ax);CHKERRQ(ierr); 35a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 36a3df083aSStefano Zampini PetscScalar sum,val; 37a3df083aSStefano Zampini const PetscInt *idxs; 38a3df083aSStefano Zampini PetscInt nz,j; 39a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 40a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 41a3df083aSStefano Zampini sum = 0.; 42a3df083aSStefano Zampini if (ctx->apply_p0) { 43a3df083aSStefano Zampini val = ctx->work[idxs[nz-1]]; 44a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 45a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 46a3df083aSStefano Zampini ctx->work[idxs[j]] += val; 47a3df083aSStefano Zampini } 48a3df083aSStefano Zampini } else { 49a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 50a3df083aSStefano Zampini sum += ctx->work[idxs[j]]; 51a3df083aSStefano Zampini } 52a3df083aSStefano Zampini } 53a3df083aSStefano Zampini ctx->work[idxs[nz-1]] -= sum; 54a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 55a3df083aSStefano Zampini } 56a3df083aSStefano Zampini ierr = VecPlaceArray(x,ctx->work);CHKERRQ(ierr); 57a3df083aSStefano Zampini reset_x = PETSC_TRUE; 58a3df083aSStefano Zampini } 59a3df083aSStefano Zampini if (transpose) { 60a3df083aSStefano Zampini ierr = MatMultTranspose(ctx->A,x,y);CHKERRQ(ierr); 61a3df083aSStefano Zampini } else { 62a3df083aSStefano Zampini ierr = MatMult(ctx->A,x,y);CHKERRQ(ierr); 63a3df083aSStefano Zampini } 64a3df083aSStefano Zampini if (reset_x) { 65a3df083aSStefano Zampini ierr = VecResetArray(x);CHKERRQ(ierr); 66a3df083aSStefano Zampini } 67a3df083aSStefano Zampini if (apply_left) { 68a3df083aSStefano Zampini PetscScalar *ay; 69a3df083aSStefano Zampini PetscInt i; 70a3df083aSStefano Zampini 71a3df083aSStefano Zampini ierr = VecGetArray(y,&ay);CHKERRQ(ierr); 72a3df083aSStefano Zampini for (i=0;i<ctx->benign_n;i++) { 73a3df083aSStefano Zampini PetscScalar sum,val; 74a3df083aSStefano Zampini const PetscInt *idxs; 75a3df083aSStefano Zampini PetscInt nz,j; 76a3df083aSStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 77a3df083aSStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 78a3df083aSStefano Zampini val = -ay[idxs[nz-1]]; 79a3df083aSStefano Zampini if (ctx->apply_p0) { 80a3df083aSStefano Zampini sum = 0.; 81a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 82a3df083aSStefano Zampini sum += ay[idxs[j]]; 83a3df083aSStefano Zampini ay[idxs[j]] += val; 84a3df083aSStefano Zampini } 85a3df083aSStefano Zampini ay[idxs[nz-1]] += sum; 86a3df083aSStefano Zampini } else { 87a3df083aSStefano Zampini for (j=0;j<nz-1;j++) { 88a3df083aSStefano Zampini ay[idxs[j]] += val; 89a3df083aSStefano Zampini } 90a3df083aSStefano Zampini ay[idxs[nz-1]] = 0.; 91a3df083aSStefano Zampini } 92a3df083aSStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);CHKERRQ(ierr); 93a3df083aSStefano Zampini } 94a3df083aSStefano Zampini ierr = VecRestoreArray(y,&ay);CHKERRQ(ierr); 95a3df083aSStefano Zampini } 96a3df083aSStefano Zampini PetscFunctionReturn(0); 97a3df083aSStefano Zampini } 98a3df083aSStefano Zampini 99a3df083aSStefano Zampini #undef __FUNCT__ 100a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMultTranspose_Private" 101a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y) 102a3df083aSStefano Zampini { 103a3df083aSStefano Zampini PetscErrorCode ierr; 104a3df083aSStefano Zampini 105a3df083aSStefano Zampini PetscFunctionBegin; 106a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);CHKERRQ(ierr); 107a3df083aSStefano Zampini PetscFunctionReturn(0); 108a3df083aSStefano Zampini } 109a3df083aSStefano Zampini 110a3df083aSStefano Zampini #undef __FUNCT__ 111a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignMatMult_Private" 112a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y) 113a3df083aSStefano Zampini { 114a3df083aSStefano Zampini PetscErrorCode ierr; 115a3df083aSStefano Zampini 116a3df083aSStefano Zampini PetscFunctionBegin; 117a3df083aSStefano Zampini ierr = PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);CHKERRQ(ierr); 118a3df083aSStefano Zampini PetscFunctionReturn(0); 119a3df083aSStefano Zampini } 120a3df083aSStefano Zampini 121a3df083aSStefano Zampini #undef __FUNCT__ 122a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignShellMat" 123a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore) 124a3df083aSStefano Zampini { 125a3df083aSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 126a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 127a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 128a3df083aSStefano Zampini PCBDDCBenignMatMult_ctx ctx; 129a3df083aSStefano Zampini PetscErrorCode ierr; 130a3df083aSStefano Zampini 131a3df083aSStefano Zampini PetscFunctionBegin; 132a3df083aSStefano Zampini if (!restore) { 133a3df083aSStefano Zampini Mat A,A_IB,A_BI; 134a3df083aSStefano Zampini PetscScalar *work; 135df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 136a3df083aSStefano Zampini 137a3df083aSStefano Zampini /* local mat */ 138a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A);CHKERRQ(ierr); 139a3df083aSStefano Zampini ierr = MatSetSizes(A,pcis->n,pcis->n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 140a3df083aSStefano Zampini ierr = MatSetType(A,MATSHELL);CHKERRQ(ierr); 141a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 142a3df083aSStefano Zampini ierr = MatShellSetOperation(A,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 143a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 144a3df083aSStefano Zampini ierr = MatShellSetContext(A,ctx);CHKERRQ(ierr); 145a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 146a3df083aSStefano Zampini ctx->apply_right = PETSC_TRUE; 147a3df083aSStefano Zampini ctx->apply_p0 = PETSC_TRUE; 148a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 149a3df083aSStefano Zampini ctx->benign_zerodiag_subs = pcbddc->benign_zerodiag_subs; 150a3df083aSStefano Zampini ctx->A = matis->A; 151a3df083aSStefano Zampini ierr = PetscMalloc1(pcis->n,&work);CHKERRQ(ierr); 152a3df083aSStefano Zampini ctx->work = work; 153a3df083aSStefano Zampini ierr = MatSetUp(A);CHKERRQ(ierr); 154a3df083aSStefano Zampini ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 155a3df083aSStefano Zampini ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 156a3df083aSStefano Zampini matis->A = A; 157a3df083aSStefano Zampini 158a3df083aSStefano Zampini /* A_IB */ 159a3df083aSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&A_IB);CHKERRQ(ierr); 160a3df083aSStefano Zampini ierr = MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 161a3df083aSStefano Zampini ierr = MatSetType(A_IB,MATSHELL);CHKERRQ(ierr); 162a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);CHKERRQ(ierr); 163a3df083aSStefano Zampini ierr = MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);CHKERRQ(ierr); 164a3df083aSStefano Zampini ierr = PetscNew(&ctx);CHKERRQ(ierr); 165a3df083aSStefano Zampini ierr = MatShellSetContext(A_IB,ctx);CHKERRQ(ierr); 166a3df083aSStefano Zampini ctx->apply_left = PETSC_TRUE; 167a3df083aSStefano Zampini ctx->apply_right = PETSC_FALSE; 168a3df083aSStefano Zampini ctx->apply_p0 = PETSC_FALSE; 169a3df083aSStefano Zampini ctx->benign_n = pcbddc->benign_n; 170059032f7SStefano Zampini if (reuse) { 171a3df083aSStefano Zampini ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs; 172059032f7SStefano Zampini } else { /* TODO: could be optimized for successive solves */ 173059032f7SStefano Zampini ISLocalToGlobalMapping N_to_D; 174059032f7SStefano Zampini PetscInt i; 175059032f7SStefano Zampini 176059032f7SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);CHKERRQ(ierr); 177059032f7SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);CHKERRQ(ierr); 178059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 179059032f7SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 180059032f7SStefano Zampini } 181059032f7SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_D);CHKERRQ(ierr); 182059032f7SStefano Zampini } 183a3df083aSStefano Zampini ctx->A = pcis->A_IB; 184a3df083aSStefano Zampini ctx->work = work; 185a3df083aSStefano Zampini ierr = MatSetUp(A_IB);CHKERRQ(ierr); 186a3df083aSStefano Zampini ierr = MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 187a3df083aSStefano Zampini ierr = MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 188a3df083aSStefano Zampini pcis->A_IB = A_IB; 189a3df083aSStefano Zampini 190a3df083aSStefano Zampini /* A_BI as A_IB^T */ 191a3df083aSStefano Zampini ierr = MatCreateTranspose(A_IB,&A_BI);CHKERRQ(ierr); 192a3df083aSStefano Zampini pcbddc->benign_original_mat = pcis->A_BI; 193a3df083aSStefano Zampini pcis->A_BI = A_BI; 194a3df083aSStefano Zampini } else { 195059032f7SStefano Zampini PCBDDCReuseSolvers reuse = pcbddc->sub_schurs->reuse_solver; 196059032f7SStefano Zampini 197a3df083aSStefano Zampini ierr = MatShellGetContext(matis->A,&ctx);CHKERRQ(ierr); 198a3df083aSStefano Zampini ierr = MatDestroy(&matis->A);CHKERRQ(ierr); 199a3df083aSStefano Zampini matis->A = ctx->A; 200a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 201a3df083aSStefano Zampini ierr = MatShellGetContext(pcis->A_IB,&ctx);CHKERRQ(ierr); 202a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_IB);CHKERRQ(ierr); 203a3df083aSStefano Zampini pcis->A_IB = ctx->A; 204059032f7SStefano Zampini if (!reuse) { 205059032f7SStefano Zampini PetscInt i; 206059032f7SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 207059032f7SStefano Zampini ierr = ISDestroy(&ctx->benign_zerodiag_subs[i]);CHKERRQ(ierr); 208059032f7SStefano Zampini } 209059032f7SStefano Zampini ierr = PetscFree(ctx->benign_zerodiag_subs);CHKERRQ(ierr); 210059032f7SStefano Zampini } 211a3df083aSStefano Zampini ierr = PetscFree(ctx->work);CHKERRQ(ierr); 212a3df083aSStefano Zampini ierr = PetscFree(ctx);CHKERRQ(ierr); 213a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_BI);CHKERRQ(ierr); 214a3df083aSStefano Zampini pcis->A_BI = pcbddc->benign_original_mat; 215a3df083aSStefano Zampini pcbddc->benign_original_mat = NULL; 216a3df083aSStefano Zampini } 217a3df083aSStefano Zampini PetscFunctionReturn(0); 218a3df083aSStefano Zampini } 219a3df083aSStefano Zampini 220a3df083aSStefano Zampini /* used just in bddc debug mode */ 221a3df083aSStefano Zampini #undef __FUNCT__ 222a3df083aSStefano Zampini #define __FUNCT__ "PCBDDCBenignProject" 223a3df083aSStefano Zampini PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B) 224a3df083aSStefano Zampini { 225a3df083aSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 226a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 227a3df083aSStefano Zampini Mat An; 228a3df083aSStefano Zampini PetscErrorCode ierr; 229a3df083aSStefano Zampini 230a3df083aSStefano Zampini PetscFunctionBegin; 231a3df083aSStefano Zampini ierr = MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);CHKERRQ(ierr); 232a3df083aSStefano Zampini ierr = MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);CHKERRQ(ierr); 233a3df083aSStefano Zampini if (is1) { 234a3df083aSStefano Zampini ierr = MatGetSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);CHKERRQ(ierr); 235a3df083aSStefano Zampini ierr = MatDestroy(&An);CHKERRQ(ierr); 236a3df083aSStefano Zampini } else { 237a3df083aSStefano Zampini *B = An; 238a3df083aSStefano Zampini } 239a3df083aSStefano Zampini PetscFunctionReturn(0); 240a3df083aSStefano Zampini } 241a3df083aSStefano Zampini 2421cf9b237SStefano Zampini /* TODO: add reuse flag */ 2431cf9b237SStefano Zampini #undef __FUNCT__ 2441cf9b237SStefano Zampini #define __FUNCT__ "MatSeqAIJCompress" 2451cf9b237SStefano Zampini PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B) 2461cf9b237SStefano Zampini { 2471cf9b237SStefano Zampini Mat Bt; 2481cf9b237SStefano Zampini PetscScalar *a,*bdata; 2491cf9b237SStefano Zampini const PetscInt *ii,*ij; 2501cf9b237SStefano Zampini PetscInt m,n,i,nnz,*bii,*bij; 2511cf9b237SStefano Zampini PetscBool flg_row; 2521cf9b237SStefano Zampini PetscErrorCode ierr; 2531cf9b237SStefano Zampini 2541cf9b237SStefano Zampini PetscFunctionBegin; 2551cf9b237SStefano Zampini ierr = MatGetSize(A,&n,&m);CHKERRQ(ierr); 2561cf9b237SStefano Zampini ierr = MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2571cf9b237SStefano Zampini ierr = MatSeqAIJGetArray(A,&a);CHKERRQ(ierr); 2581cf9b237SStefano Zampini nnz = n; 2591cf9b237SStefano Zampini for (i=0;i<ii[n];i++) { 2601cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++; 2611cf9b237SStefano Zampini } 2621cf9b237SStefano Zampini ierr = PetscMalloc1(n+1,&bii);CHKERRQ(ierr); 2631cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bij);CHKERRQ(ierr); 2641cf9b237SStefano Zampini ierr = PetscMalloc1(nnz,&bdata);CHKERRQ(ierr); 2651cf9b237SStefano Zampini nnz = 0; 2661cf9b237SStefano Zampini bii[0] = 0; 2671cf9b237SStefano Zampini for (i=0;i<n;i++) { 2681cf9b237SStefano Zampini PetscInt j; 2691cf9b237SStefano Zampini for (j=ii[i];j<ii[i+1];j++) { 2701cf9b237SStefano Zampini PetscScalar entry = a[j]; 2711cf9b237SStefano Zampini if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || ij[j] == i) { 2721cf9b237SStefano Zampini bij[nnz] = ij[j]; 2731cf9b237SStefano Zampini bdata[nnz] = entry; 2741cf9b237SStefano Zampini nnz++; 2751cf9b237SStefano Zampini } 2761cf9b237SStefano Zampini } 2771cf9b237SStefano Zampini bii[i+1] = nnz; 2781cf9b237SStefano Zampini } 2791cf9b237SStefano Zampini ierr = MatSeqAIJRestoreArray(A,&a);CHKERRQ(ierr); 2801cf9b237SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);CHKERRQ(ierr); 2811cf9b237SStefano Zampini ierr = MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);CHKERRQ(ierr); 2821cf9b237SStefano Zampini { 2831cf9b237SStefano Zampini Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data); 2841cf9b237SStefano Zampini b->free_a = PETSC_TRUE; 2851cf9b237SStefano Zampini b->free_ij = PETSC_TRUE; 2861cf9b237SStefano Zampini } 2871cf9b237SStefano Zampini *B = Bt; 2881cf9b237SStefano Zampini PetscFunctionReturn(0); 2891cf9b237SStefano Zampini } 2901cf9b237SStefano Zampini 291674ae819SStefano Zampini #undef __FUNCT__ 2924f1b2e48SStefano Zampini #define __FUNCT__ "MatDetectDisconnectedComponents" 2934f1b2e48SStefano Zampini PetscErrorCode MatDetectDisconnectedComponents(Mat A, PetscBool filter, PetscInt *ncc, IS* cc[]) 2944f1b2e48SStefano Zampini { 2954f1b2e48SStefano Zampini Mat B; 2964f1b2e48SStefano Zampini IS is_dummy,*cc_n; 2974f1b2e48SStefano Zampini ISLocalToGlobalMapping l2gmap_dummy; 2984f1b2e48SStefano Zampini PCBDDCGraph graph; 2994f1b2e48SStefano Zampini PetscInt i,n; 3004f1b2e48SStefano Zampini PetscInt *xadj,*adjncy; 3014f1b2e48SStefano Zampini PetscInt *xadj_filtered,*adjncy_filtered; 3024f1b2e48SStefano Zampini PetscBool flg_row,isseqaij; 3034f1b2e48SStefano Zampini PetscErrorCode ierr; 3044f1b2e48SStefano Zampini 3054f1b2e48SStefano Zampini PetscFunctionBegin; 30663c961adSStefano Zampini if (!A->rmap->N || !A->cmap->N) { 30763c961adSStefano Zampini *ncc = 0; 30863c961adSStefano Zampini *cc = NULL; 30963c961adSStefano Zampini PetscFunctionReturn(0); 31063c961adSStefano Zampini } 3114f1b2e48SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 3124f1b2e48SStefano Zampini if (!isseqaij && filter) { 3131cf9b237SStefano Zampini PetscBool isseqdense; 3141cf9b237SStefano Zampini 3151cf9b237SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);CHKERRQ(ierr); 3161cf9b237SStefano Zampini if (!isseqdense) { 3174f1b2e48SStefano Zampini ierr = MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); 3181cf9b237SStefano Zampini } else { /* TODO: rectangular case and LDA */ 3191cf9b237SStefano Zampini PetscScalar *array; 3201cf9b237SStefano Zampini PetscReal chop=1.e-6; 3211cf9b237SStefano Zampini 3221cf9b237SStefano Zampini ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 3231cf9b237SStefano Zampini ierr = MatDenseGetArray(B,&array);CHKERRQ(ierr); 3241cf9b237SStefano Zampini ierr = MatGetSize(B,&n,NULL);CHKERRQ(ierr); 3251cf9b237SStefano Zampini for (i=0;i<n;i++) { 3261cf9b237SStefano Zampini PetscInt j; 3271cf9b237SStefano Zampini for (j=i+1;j<n;j++) { 3281cf9b237SStefano Zampini PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)])); 3291cf9b237SStefano Zampini if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.; 3301cf9b237SStefano Zampini if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.; 3311cf9b237SStefano Zampini } 3321cf9b237SStefano Zampini } 3331cf9b237SStefano Zampini ierr = MatDenseRestoreArray(B,&array);CHKERRQ(ierr); 3341cf9b237SStefano Zampini ierr = MatConvert(B,MATSEQAIJ,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); 3351cf9b237SStefano Zampini } 3364f1b2e48SStefano Zampini } else { 3374f1b2e48SStefano Zampini B = A; 3384f1b2e48SStefano Zampini } 3394f1b2e48SStefano Zampini ierr = MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3404f1b2e48SStefano Zampini 3414f1b2e48SStefano Zampini /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */ 3424f1b2e48SStefano Zampini if (filter) { 3434f1b2e48SStefano Zampini PetscScalar *data; 3444f1b2e48SStefano Zampini PetscInt j,cum; 3454f1b2e48SStefano Zampini 3464f1b2e48SStefano Zampini ierr = PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);CHKERRQ(ierr); 3474f1b2e48SStefano Zampini ierr = MatSeqAIJGetArray(B,&data);CHKERRQ(ierr); 3484f1b2e48SStefano Zampini cum = 0; 3494f1b2e48SStefano Zampini for (i=0;i<n;i++) { 3504f1b2e48SStefano Zampini PetscInt t; 3514f1b2e48SStefano Zampini 3524f1b2e48SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) { 3534f1b2e48SStefano Zampini if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) { 3544f1b2e48SStefano Zampini continue; 3554f1b2e48SStefano Zampini } 3564f1b2e48SStefano Zampini adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j]; 3574f1b2e48SStefano Zampini } 3584f1b2e48SStefano Zampini t = xadj_filtered[i]; 3594f1b2e48SStefano Zampini xadj_filtered[i] = cum; 3604f1b2e48SStefano Zampini cum += t; 3614f1b2e48SStefano Zampini } 3624f1b2e48SStefano Zampini ierr = MatSeqAIJRestoreArray(B,&data);CHKERRQ(ierr); 3634f1b2e48SStefano Zampini } else { 3644f1b2e48SStefano Zampini xadj_filtered = NULL; 3654f1b2e48SStefano Zampini adjncy_filtered = NULL; 3664f1b2e48SStefano Zampini } 3674f1b2e48SStefano Zampini 3684f1b2e48SStefano Zampini /* compute local connected components using PCBDDCGraph */ 3694f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);CHKERRQ(ierr); 3704f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);CHKERRQ(ierr); 3714f1b2e48SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 3724f1b2e48SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 3734f1b2e48SStefano Zampini ierr = PCBDDCGraphInit(graph,l2gmap_dummy,n);CHKERRQ(ierr); 3744f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_dummy);CHKERRQ(ierr); 3754f1b2e48SStefano Zampini if (xadj_filtered) { 3764f1b2e48SStefano Zampini graph->xadj = xadj_filtered; 3774f1b2e48SStefano Zampini graph->adjncy = adjncy_filtered; 3784f1b2e48SStefano Zampini } else { 3794f1b2e48SStefano Zampini graph->xadj = xadj; 3804f1b2e48SStefano Zampini graph->adjncy = adjncy; 3814f1b2e48SStefano Zampini } 3824f1b2e48SStefano Zampini ierr = PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);CHKERRQ(ierr); 3834f1b2e48SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 3844f1b2e48SStefano Zampini /* partial clean up */ 3854f1b2e48SStefano Zampini ierr = PetscFree2(xadj_filtered,adjncy_filtered);CHKERRQ(ierr); 3864f1b2e48SStefano Zampini ierr = MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 3871cf9b237SStefano Zampini if (A != B) { 3884f1b2e48SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 3894f1b2e48SStefano Zampini } 3904f1b2e48SStefano Zampini 3914f1b2e48SStefano Zampini /* get back data */ 3921cf9b237SStefano Zampini if (ncc) *ncc = graph->ncc; 3931cf9b237SStefano Zampini if (cc) { 3944f1b2e48SStefano Zampini ierr = PetscMalloc1(graph->ncc,&cc_n);CHKERRQ(ierr); 3954f1b2e48SStefano Zampini for (i=0;i<graph->ncc;i++) { 3964f1b2e48SStefano 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); 3974f1b2e48SStefano Zampini } 3984f1b2e48SStefano Zampini *cc = cc_n; 3991cf9b237SStefano Zampini } 4004f1b2e48SStefano Zampini /* clean up graph */ 4014f1b2e48SStefano Zampini graph->xadj = 0; 4024f1b2e48SStefano Zampini graph->adjncy = 0; 4034f1b2e48SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 4044f1b2e48SStefano Zampini PetscFunctionReturn(0); 4054f1b2e48SStefano Zampini } 4064f1b2e48SStefano Zampini 4074f1b2e48SStefano Zampini #undef __FUNCT__ 4085408967cSStefano Zampini #define __FUNCT__ "PCBDDCBenignCheck" 4095408967cSStefano Zampini PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag) 4105408967cSStefano Zampini { 4115408967cSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 4125408967cSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 413dee84bffSStefano Zampini IS dirIS = NULL; 4144f1b2e48SStefano Zampini PetscInt i; 4155408967cSStefano Zampini PetscErrorCode ierr; 4165408967cSStefano Zampini 4175408967cSStefano Zampini PetscFunctionBegin; 418dee84bffSStefano Zampini ierr = PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);CHKERRQ(ierr); 4195408967cSStefano Zampini if (zerodiag) { 4205408967cSStefano Zampini Mat A; 4215408967cSStefano Zampini Vec vec3_N; 4225408967cSStefano Zampini PetscScalar *vals; 4235408967cSStefano Zampini const PetscInt *idxs; 424d12d3064SStefano Zampini PetscInt nz,*count; 4255408967cSStefano Zampini 4265408967cSStefano Zampini /* p0 */ 4275408967cSStefano Zampini ierr = VecSet(pcis->vec1_N,0.);CHKERRQ(ierr); 4285408967cSStefano Zampini ierr = PetscMalloc1(pcis->n,&vals);CHKERRQ(ierr); 4295408967cSStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 4305408967cSStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 4314f1b2e48SStefano Zampini for (i=0;i<nz;i++) vals[i] = 1.; 4325408967cSStefano Zampini ierr = VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4335408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 4345408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 4355408967cSStefano Zampini /* v_I */ 4365408967cSStefano Zampini ierr = VecSetRandom(pcis->vec2_N,NULL);CHKERRQ(ierr); 4375408967cSStefano Zampini for (i=0;i<nz;i++) vals[i] = 0.; 4385408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4395408967cSStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 4405408967cSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4415408967cSStefano Zampini for (i=0;i<pcis->n_B;i++) vals[i] = 0.; 4425408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4435408967cSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 4445408967cSStefano Zampini if (dirIS) { 4455408967cSStefano Zampini PetscInt n; 4465408967cSStefano Zampini 4475408967cSStefano Zampini ierr = ISGetLocalSize(dirIS,&n);CHKERRQ(ierr); 4485408967cSStefano Zampini ierr = ISGetIndices(dirIS,&idxs);CHKERRQ(ierr); 4495408967cSStefano Zampini for (i=0;i<n;i++) vals[i] = 0.; 4505408967cSStefano Zampini ierr = VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 4515408967cSStefano Zampini ierr = ISRestoreIndices(dirIS,&idxs);CHKERRQ(ierr); 4525408967cSStefano Zampini } 4535408967cSStefano Zampini ierr = VecAssemblyBegin(pcis->vec2_N);CHKERRQ(ierr); 4545408967cSStefano Zampini ierr = VecAssemblyEnd(pcis->vec2_N);CHKERRQ(ierr); 4555408967cSStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&vec3_N);CHKERRQ(ierr); 4565408967cSStefano Zampini ierr = VecSet(vec3_N,0.);CHKERRQ(ierr); 4575408967cSStefano Zampini ierr = MatISGetLocalMat(pc->mat,&A);CHKERRQ(ierr); 4585408967cSStefano Zampini ierr = MatMult(A,pcis->vec1_N,vec3_N);CHKERRQ(ierr); 4595408967cSStefano Zampini ierr = VecDot(vec3_N,pcis->vec2_N,&vals[0]);CHKERRQ(ierr); 4604f1b2e48SStefano Zampini if (PetscAbsScalar(vals[0]) > PETSC_SMALL) { 4615408967cSStefano 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])); 4625408967cSStefano Zampini } 4635408967cSStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4645408967cSStefano Zampini ierr = VecDestroy(&vec3_N);CHKERRQ(ierr); 465d12d3064SStefano Zampini 466d12d3064SStefano Zampini /* there should not be any pressure dofs lying on the interface */ 467d12d3064SStefano Zampini ierr = PetscCalloc1(pcis->n,&count);CHKERRQ(ierr); 468d12d3064SStefano Zampini ierr = ISGetIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 469d12d3064SStefano Zampini for (i=0;i<pcis->n_B;i++) count[idxs[i]]++; 470d12d3064SStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,&idxs);CHKERRQ(ierr); 471d12d3064SStefano Zampini ierr = ISGetIndices(zerodiag,&idxs);CHKERRQ(ierr); 472d4d8cf7bSStefano Zampini for (i=0;i<nz;i++) { 473d12d3064SStefano Zampini if (count[idxs[i]]) { 474d12d3064SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %d is an interface dof",idxs[i]); 475d12d3064SStefano Zampini } 476d4d8cf7bSStefano Zampini } 477d12d3064SStefano Zampini ierr = ISRestoreIndices(zerodiag,&idxs);CHKERRQ(ierr); 478d12d3064SStefano Zampini ierr = PetscFree(count);CHKERRQ(ierr); 4795408967cSStefano Zampini } 480dee84bffSStefano Zampini ierr = ISDestroy(&dirIS);CHKERRQ(ierr); 4815408967cSStefano Zampini 4825408967cSStefano Zampini /* check PCBDDCBenignGetOrSetP0 */ 4835408967cSStefano Zampini ierr = VecSetRandom(pcis->vec1_global,NULL);CHKERRQ(ierr); 4844f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i; 4855408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);CHKERRQ(ierr); 4864f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1; 4875408967cSStefano Zampini ierr = PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);CHKERRQ(ierr); 4884f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 4894f1b2e48SStefano Zampini if ((PetscInt)PetscRealPart(pcbddc->benign_p0[i]) != -PetscGlobalRank-i) { 4904f1b2e48SStefano 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); 4914f1b2e48SStefano Zampini } 4925408967cSStefano Zampini } 4935408967cSStefano Zampini PetscFunctionReturn(0); 4945408967cSStefano Zampini } 4955408967cSStefano Zampini 4965408967cSStefano Zampini #undef __FUNCT__ 497339f8db1SStefano Zampini #define __FUNCT__ "PCBDDCBenignDetectSaddlePoint" 498339f8db1SStefano Zampini PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, IS *zerodiaglocal) 499339f8db1SStefano Zampini { 500339f8db1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5014f1b2e48SStefano Zampini IS pressures,zerodiag,*zerodiag_subs; 502b0f5fe93SStefano Zampini PetscInt nz,n; 5034f1b2e48SStefano Zampini PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag; 504339f8db1SStefano Zampini PetscErrorCode ierr; 505339f8db1SStefano Zampini 506339f8db1SStefano Zampini PetscFunctionBegin; 5079f47a83aSStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 5089f47a83aSStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 509339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 510a3df083aSStefano Zampini for (n=0;n<pcbddc->benign_n;n++) { 511a3df083aSStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[n]);CHKERRQ(ierr); 512a3df083aSStefano Zampini } 513a3df083aSStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 514a3df083aSStefano Zampini pcbddc->benign_n = 0; 5154f1b2e48SStefano Zampini /* if a local info on dofs is present, assumes the last field is represented by "pressures" 5164f1b2e48SStefano Zampini otherwise, it uses only zerodiagonal dofs (ok if the pressure block is all zero; it could fail if it is not) 5174f1b2e48SStefano Zampini Checks if all the pressure dofs in each subdomain have a zero diagonal 5184f1b2e48SStefano Zampini If not, a change of basis on pressures is not needed 5194f1b2e48SStefano Zampini since the local Schur complements are SPD 5204f1b2e48SStefano Zampini */ 5214f1b2e48SStefano Zampini has_null_pressures = PETSC_TRUE; 5224f1b2e48SStefano Zampini have_null = PETSC_TRUE; 52340fa8d13SStefano Zampini if (pcbddc->n_ISForDofsLocal) { 5244f1b2e48SStefano Zampini PetscInt npl,*idxs,p = pcbddc->n_ISForDofsLocal-1; 5254f1b2e48SStefano Zampini 5264f1b2e48SStefano Zampini /* Dofs splitting for BDDC cannot have PETSC_COMM_SELF, so create a sequential IS */ 5274f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[p],&npl);CHKERRQ(ierr); 5284f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 5294f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,npl,idxs,PETSC_COPY_VALUES,&pressures);CHKERRQ(ierr); 530ba14f8e3SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[p],(const PetscInt**)&idxs);CHKERRQ(ierr); 53140fa8d13SStefano Zampini ierr = ISSorted(pressures,&sorted);CHKERRQ(ierr); 53240fa8d13SStefano Zampini if (!sorted) { 53340fa8d13SStefano Zampini ierr = ISSort(pressures);CHKERRQ(ierr); 53440fa8d13SStefano Zampini } 53540fa8d13SStefano Zampini } else { 53640fa8d13SStefano Zampini pressures = NULL; 53740fa8d13SStefano Zampini } 53897d764eeSStefano Zampini /* pcis has not been setup yet, so get the local size from the subdomain matrix */ 53997d764eeSStefano Zampini ierr = MatGetLocalSize(pcbddc->local_mat,&n,NULL);CHKERRQ(ierr); 54097d764eeSStefano Zampini ierr = MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);CHKERRQ(ierr); 541339f8db1SStefano Zampini ierr = ISSorted(zerodiag,&sorted);CHKERRQ(ierr); 542339f8db1SStefano Zampini if (!sorted) { 543339f8db1SStefano Zampini ierr = ISSort(zerodiag);CHKERRQ(ierr); 544339f8db1SStefano Zampini } 545339f8db1SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 5464f1b2e48SStefano Zampini if (!nz) { 5474f1b2e48SStefano Zampini if (n) have_null = PETSC_FALSE; 5484f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 54940fa8d13SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 55040fa8d13SStefano Zampini } 5514f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 5524f1b2e48SStefano Zampini /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */ 5534f1b2e48SStefano Zampini zerodiag_subs = NULL; 5544f1b2e48SStefano Zampini pcbddc->benign_n = 0; 5554f1b2e48SStefano Zampini if (has_null_pressures) { 5564f1b2e48SStefano Zampini IS *subs; 5574f1b2e48SStefano Zampini PetscInt nsubs,i; 5584f1b2e48SStefano Zampini 5594f1b2e48SStefano Zampini subs = pcbddc->local_subs; 5604f1b2e48SStefano Zampini nsubs = pcbddc->n_local_subs; 5614f1b2e48SStefano Zampini if (nsubs > 1) { 5624f1b2e48SStefano Zampini ierr = PetscCalloc1(nsubs,&zerodiag_subs);CHKERRQ(ierr); 5634f1b2e48SStefano Zampini for (i=0;i<nsubs;i++) { 5644f1b2e48SStefano Zampini ISLocalToGlobalMapping l2g; 5654f1b2e48SStefano Zampini IS t_zerodiag_subs; 5664f1b2e48SStefano Zampini PetscInt nl; 5674f1b2e48SStefano Zampini 5684f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(subs[i],&l2g);CHKERRQ(ierr); 5694f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,zerodiag,&t_zerodiag_subs);CHKERRQ(ierr); 5704f1b2e48SStefano Zampini ierr = ISGetLocalSize(t_zerodiag_subs,&nl);CHKERRQ(ierr); 5714f1b2e48SStefano Zampini if (nl) { 5724f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5734f1b2e48SStefano Zampini 5744f1b2e48SStefano Zampini if (pressures) { 5754f1b2e48SStefano Zampini IS t_pressure_subs; 5764f1b2e48SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);CHKERRQ(ierr); 5774f1b2e48SStefano Zampini ierr = ISEqual(t_pressure_subs,t_zerodiag_subs,&valid);CHKERRQ(ierr); 5784f1b2e48SStefano Zampini ierr = ISDestroy(&t_pressure_subs);CHKERRQ(ierr); 5794f1b2e48SStefano Zampini } 5804f1b2e48SStefano Zampini if (valid) { 5814f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[pcbddc->benign_n]);CHKERRQ(ierr); 5824f1b2e48SStefano Zampini pcbddc->benign_n++; 5834f1b2e48SStefano Zampini } else { 5844f1b2e48SStefano Zampini recompute_zerodiag = PETSC_TRUE; 5854f1b2e48SStefano Zampini } 5864f1b2e48SStefano Zampini } 5874f1b2e48SStefano Zampini ierr = ISDestroy(&t_zerodiag_subs);CHKERRQ(ierr); 5884f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2g);CHKERRQ(ierr); 5894f1b2e48SStefano Zampini } 5904f1b2e48SStefano Zampini } else { /* there's just one subdomain (or zero if they have not been detected */ 5914f1b2e48SStefano Zampini PetscBool valid = PETSC_TRUE; 5924f1b2e48SStefano Zampini if (pressures) { 5934f1b2e48SStefano Zampini ierr = ISEqual(pressures,zerodiag,&valid);CHKERRQ(ierr); 5944f1b2e48SStefano Zampini } 5954f1b2e48SStefano Zampini if (valid) { 5964f1b2e48SStefano Zampini pcbddc->benign_n = 1; 597ca92afb2SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&zerodiag_subs);CHKERRQ(ierr); 5984f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag);CHKERRQ(ierr); 5994f1b2e48SStefano Zampini zerodiag_subs[0] = zerodiag; 6004f1b2e48SStefano Zampini } 6014f1b2e48SStefano Zampini } 6024f1b2e48SStefano Zampini } 6034f1b2e48SStefano Zampini 6044f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 6054f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6064f1b2e48SStefano Zampini recompute_zerodiag = PETSC_FALSE; 6074f1b2e48SStefano Zampini has_null_pressures = PETSC_FALSE; 6084f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6094f1b2e48SStefano Zampini } 6104f1b2e48SStefano Zampini 6114f1b2e48SStefano Zampini /* final check for null pressures */ 6124f1b2e48SStefano Zampini if (zerodiag && pressures) { 6134f1b2e48SStefano Zampini PetscInt nz,np; 6144f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 6154f1b2e48SStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 6164f1b2e48SStefano Zampini if (nz != np) have_null = PETSC_FALSE; 6174f1b2e48SStefano Zampini } 6184f1b2e48SStefano Zampini 6194f1b2e48SStefano Zampini if (recompute_zerodiag) { 6204f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiag);CHKERRQ(ierr); 6214f1b2e48SStefano Zampini if (pcbddc->benign_n == 1) { 6224f1b2e48SStefano Zampini ierr = PetscObjectReference((PetscObject)zerodiag_subs[0]);CHKERRQ(ierr); 6234f1b2e48SStefano Zampini zerodiag = zerodiag_subs[0]; 6244f1b2e48SStefano Zampini } else { 6254f1b2e48SStefano Zampini PetscInt i,nzn,*new_idxs; 6264f1b2e48SStefano Zampini 6274f1b2e48SStefano Zampini nzn = 0; 6284f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6294f1b2e48SStefano Zampini PetscInt ns; 6304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6314f1b2e48SStefano Zampini nzn += ns; 6324f1b2e48SStefano Zampini } 6334f1b2e48SStefano Zampini ierr = PetscMalloc1(nzn,&new_idxs);CHKERRQ(ierr); 6344f1b2e48SStefano Zampini nzn = 0; 6354f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6364f1b2e48SStefano Zampini PetscInt ns,*idxs; 6374f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&ns);CHKERRQ(ierr); 6384f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6394f1b2e48SStefano Zampini ierr = PetscMemcpy(new_idxs+nzn,idxs,ns*sizeof(PetscInt));CHKERRQ(ierr); 6404f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 6414f1b2e48SStefano Zampini nzn += ns; 6424f1b2e48SStefano Zampini } 6434f1b2e48SStefano Zampini ierr = PetscSortInt(nzn,new_idxs);CHKERRQ(ierr); 6444f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);CHKERRQ(ierr); 6454f1b2e48SStefano Zampini } 6464f1b2e48SStefano Zampini have_null = PETSC_FALSE; 6474f1b2e48SStefano Zampini } 6484f1b2e48SStefano Zampini 649b3afcdbeSStefano Zampini /* no-net-flux */ 650b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 651b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 652b3afcdbeSStefano Zampini MatNullSpace near_null_space; 653b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 654b3afcdbeSStefano Zampini Vec quad_vec; 655b3afcdbeSStefano Zampini PetscScalar *pvals; 656b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 657b3afcdbeSStefano Zampini 658b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 659b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 660b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 661b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 662b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 663b3afcdbeSStefano Zampini 664b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 665b3afcdbeSStefano Zampini np = 0; 666b3afcdbeSStefano Zampini if (zerodiag) { 667b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 668b3afcdbeSStefano Zampini } else if (pressures) { 669b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 670b3afcdbeSStefano Zampini } 671b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 672b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 673b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 674b3afcdbeSStefano Zampini if (np) { 675b3afcdbeSStefano Zampini const PetscInt *pidxs; 676b3afcdbeSStefano Zampini 677b3afcdbeSStefano Zampini if (zerodiag) { 678b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 679b3afcdbeSStefano Zampini } else if (pressures) { 680b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 681b3afcdbeSStefano Zampini } 682b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 683b3afcdbeSStefano Zampini if (zerodiag) { 684b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 685b3afcdbeSStefano Zampini } else if (pressures) { 686b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 687b3afcdbeSStefano Zampini } 688b3afcdbeSStefano Zampini } 689b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 690b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 691b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 692b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 693b3afcdbeSStefano Zampini 694b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 695b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 696b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = -1; 697b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 698b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 699b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 700b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 701b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 702b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 703b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) { 704b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 705b3afcdbeSStefano Zampini pvals[i] = 0.; 706b3afcdbeSStefano Zampini } 707b3afcdbeSStefano Zampini } 708b3afcdbeSStefano Zampini ierr = PetscMalloc1(matis->sf_nleaves,&dummyins);CHKERRQ(ierr); 709b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) dummyins[i] = i; 710b3afcdbeSStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf_nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 711b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 712b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 713b3afcdbeSStefano Zampini 714b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 715b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 716b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 717b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 718b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 719b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 720b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 721b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 722b3afcdbeSStefano Zampini } 723b3afcdbeSStefano Zampini 724b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7254f1b2e48SStefano Zampini if (has_null_pressures) { 7264f1b2e48SStefano Zampini IS zerodiagc; 7274f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7284f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7294f1b2e48SStefano Zampini 7304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 731339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 732339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 733339f8db1SStefano Zampini /* local change of basis for pressures */ 734339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 73597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 736339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 737339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 738339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7394f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7414f1b2e48SStefano Zampini PetscInt nzs,j; 7424f1b2e48SStefano Zampini 7434f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7444f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7454f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7464f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7474f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7484f1b2e48SStefano Zampini } 749339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 750339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 751339f8db1SStefano Zampini /* set identity on velocities */ 752339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 753339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 754339f8db1SStefano Zampini } 7554f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7564f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7579f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7584f1b2e48SStefano 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); 759339f8db1SStefano Zampini /* set change on pressures */ 7604f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7614f1b2e48SStefano Zampini PetscScalar *array; 7624f1b2e48SStefano Zampini PetscInt nzs; 7634f1b2e48SStefano Zampini 7644f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7654f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7664f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 767339f8db1SStefano Zampini PetscScalar vals[2]; 768339f8db1SStefano Zampini PetscInt cols[2]; 769339f8db1SStefano Zampini 770339f8db1SStefano Zampini cols[0] = idxs[i]; 7714f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 772339f8db1SStefano Zampini vals[0] = 1.; 773b0f5fe93SStefano Zampini vals[1] = 1.; 7744f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 775339f8db1SStefano Zampini } 7764f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7774f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7784f1b2e48SStefano Zampini array[nzs-1] = 1.; 7794f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7804f1b2e48SStefano Zampini /* store local idxs for p0 */ 7814f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7824f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 783339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7844f1b2e48SStefano Zampini } 785339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 786339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 787a3df083aSStefano Zampini /* project if needed */ 788a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 78997d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 790339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 79197d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 79297d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 793a3df083aSStefano Zampini } 7944f1b2e48SStefano Zampini /* store global idxs for p0 */ 7954f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 796339f8db1SStefano Zampini } 797ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7984f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 799b0f5fe93SStefano Zampini 800b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 801b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 802339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 803339f8db1SStefano Zampini PetscFunctionReturn(0); 804339f8db1SStefano Zampini } 805339f8db1SStefano Zampini 806339f8db1SStefano Zampini #undef __FUNCT__ 807015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 808015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 809efc2fbd9SStefano Zampini { 810efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 811efc2fbd9SStefano Zampini PetscErrorCode ierr; 812efc2fbd9SStefano Zampini 813efc2fbd9SStefano Zampini PetscFunctionBegin; 814efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 815efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 8164f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 817efc2fbd9SStefano Zampini } 818015636ebSStefano Zampini if (get) { /* use SF to get values */ 819efc2fbd9SStefano Zampini PetscScalar *array; 820efc2fbd9SStefano Zampini 821efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8224f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8234f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 824efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8254f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 8264f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 827efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 828efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 829efc2fbd9SStefano Zampini } 830efc2fbd9SStefano Zampini PetscFunctionReturn(0); 831efc2fbd9SStefano Zampini } 832efc2fbd9SStefano Zampini 833efc2fbd9SStefano Zampini #undef __FUNCT__ 834c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 835c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 836c263805aSStefano Zampini { 837c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 838c263805aSStefano Zampini PetscErrorCode ierr; 839c263805aSStefano Zampini 840c263805aSStefano Zampini PetscFunctionBegin; 841c263805aSStefano Zampini /* TODO: add error checking 842c263805aSStefano Zampini - avoid nested pop (or push) calls. 843c263805aSStefano Zampini - cannot push before pop. 8441c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 845c263805aSStefano Zampini */ 8464f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 847efc2fbd9SStefano Zampini PetscFunctionReturn(0); 848efc2fbd9SStefano Zampini } 849c263805aSStefano Zampini if (pop) { 850a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8514f1b2e48SStefano Zampini IS is_p0; 8524f1b2e48SStefano Zampini MatReuse reuse; 853c263805aSStefano Zampini 854c263805aSStefano Zampini /* extract B_0 */ 8554f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8564f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8574f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8584f1b2e48SStefano Zampini } 8594f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8604f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 861c263805aSStefano Zampini /* remove rows and cols from local problem */ 862c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 86397d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8644f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8654f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 866a3df083aSStefano Zampini } else { 867a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 868a3df083aSStefano Zampini PetscScalar *vals; 869a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 870a3df083aSStefano Zampini 871a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 872a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 873a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8740b5adadeSStefano Zampini PetscInt *nnz; 875a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 876a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 877a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 878331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 879331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 880331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 881331e053bSStefano Zampini nnz[i] = n - nnz[i]; 882331e053bSStefano Zampini } 883331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 884331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 885331e053bSStefano Zampini } 886a3df083aSStefano Zampini 887a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 888a3df083aSStefano Zampini PetscScalar *array; 889a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 890a3df083aSStefano Zampini 891a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 892a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 893a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 894a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 895a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 896a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 897a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 898a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 899a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 900a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 901a3df083aSStefano Zampini cum = 0; 902a3df083aSStefano Zampini for (j=0;j<n;j++) { 903a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 904a3df083aSStefano Zampini vals[cum] = array[j]; 905a3df083aSStefano Zampini idxs_ins[cum] = j; 906a3df083aSStefano Zampini cum++; 907a3df083aSStefano Zampini } 908a3df083aSStefano Zampini } 909a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 910a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 911a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 912a3df083aSStefano Zampini } 913a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 914a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 915a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 916a3df083aSStefano Zampini } 917c263805aSStefano Zampini } else { /* push */ 918a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9194f1b2e48SStefano Zampini PetscInt i; 9204f1b2e48SStefano Zampini 9214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9224f1b2e48SStefano Zampini PetscScalar *B0_vals; 9234f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9244f1b2e48SStefano Zampini 9254f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9264f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9277b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9284f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9294f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9304f1b2e48SStefano Zampini } 931c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 932c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 933a3df083aSStefano Zampini } else { 934a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 935a3df083aSStefano Zampini } 936c263805aSStefano Zampini } 937c263805aSStefano Zampini PetscFunctionReturn(0); 938c263805aSStefano Zampini } 939c263805aSStefano Zampini 940c263805aSStefano Zampini #undef __FUNCT__ 941b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 94208122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 943b1b3d7a2SStefano Zampini { 944b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 94508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 94608122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 94708122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 94808122e43SStefano Zampini PetscScalar *work,lwork; 94908122e43SStefano Zampini PetscScalar *St,*S,*eigv; 95008122e43SStefano Zampini PetscScalar *Sarray,*Starray; 95108122e43SStefano Zampini PetscReal *eigs,thresh; 9521b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 953f6f667cfSStefano Zampini PetscBool allocated_S_St; 95408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 95508122e43SStefano Zampini PetscReal *rwork; 95608122e43SStefano Zampini #endif 957b1b3d7a2SStefano Zampini PetscErrorCode ierr; 958b1b3d7a2SStefano Zampini 959b1b3d7a2SStefano Zampini PetscFunctionBegin; 960df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 961df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 96208122e43SStefano Zampini } 96308122e43SStefano Zampini 96406a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 96506a4e24aSStefano 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); 96606a4e24aSStefano Zampini } 96706a4e24aSStefano Zampini 968fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 969fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 970fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 971fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9721575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 973fd14bc51SStefano Zampini } 974fd14bc51SStefano Zampini 975e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 976e496cd5dSStefano 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); 977e496cd5dSStefano Zampini } 978e496cd5dSStefano Zampini 97908122e43SStefano Zampini /* max size of subsets */ 98008122e43SStefano Zampini mss = 0; 98108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 98208122e43SStefano Zampini PetscInt subset_size; 983862806e4SStefano Zampini 98408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 98508122e43SStefano Zampini mss = PetscMax(mss,subset_size); 98608122e43SStefano Zampini } 98708122e43SStefano Zampini 98808122e43SStefano Zampini /* min/max and threshold */ 98908122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 990f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 99108122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 992f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 993f6f667cfSStefano Zampini if (nmin) { 994f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 995f6f667cfSStefano Zampini } 99608122e43SStefano Zampini 99708122e43SStefano Zampini /* allocate lapack workspace */ 99808122e43SStefano Zampini cum = cum2 = 0; 99908122e43SStefano Zampini maxneigs = 0; 100008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 100108122e43SStefano Zampini PetscInt n,subset_size; 1002f6f667cfSStefano Zampini 100308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 100408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 10059162d606SStefano Zampini cum += subset_size; 10069162d606SStefano Zampini cum2 += subset_size*n; 100708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 100808122e43SStefano Zampini } 100908122e43SStefano Zampini if (mss) { 10109ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 101108122e43SStefano Zampini PetscBLASInt B_itype = 1; 101208122e43SStefano Zampini PetscBLASInt B_N = mss; 10134c6709b3SStefano Zampini PetscReal zero = 0.0; 10144c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 101508122e43SStefano Zampini 101608122e43SStefano Zampini B_lwork = -1; 101708122e43SStefano Zampini S = NULL; 101808122e43SStefano Zampini St = NULL; 1019a58a30b4SStefano Zampini eigs = NULL; 1020a58a30b4SStefano Zampini eigv = NULL; 1021a58a30b4SStefano Zampini B_iwork = NULL; 1022a58a30b4SStefano Zampini B_ifail = NULL; 1023d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1024d1710679SStefano Zampini rwork = NULL; 1025d1710679SStefano Zampini #endif 10268bec7fa6SStefano Zampini thresh = 1.0; 102708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 102808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 102908122e43SStefano 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)); 103008122e43SStefano Zampini #else 103108122e43SStefano 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)); 103208122e43SStefano Zampini #endif 103308122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 103408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 103508122e43SStefano Zampini } else { 103608122e43SStefano Zampini /* TODO */ 103708122e43SStefano Zampini } 103808122e43SStefano Zampini } else { 103908122e43SStefano Zampini lwork = 0; 104008122e43SStefano Zampini } 104108122e43SStefano Zampini 104208122e43SStefano Zampini nv = 0; 1043d62866d3SStefano 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) */ 1044d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 104508122e43SStefano Zampini } 10464c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1047f6f667cfSStefano Zampini if (allocated_S_St) { 1048f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1049f6f667cfSStefano Zampini } 1050f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 105108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 105208122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 105308122e43SStefano Zampini #endif 10549162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10559162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10569162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 105708122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10589162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 105908122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 106008122e43SStefano Zampini 106108122e43SStefano Zampini maxneigs = 0; 106272b8c272SStefano Zampini cum = cumarray = 0; 10639162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10649162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1065d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 106608122e43SStefano Zampini const PetscInt *idxs; 106708122e43SStefano Zampini 1068d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 106908122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 107008122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 107108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 107208122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10739162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10749162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 107508122e43SStefano Zampini } 1076d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 107708122e43SStefano Zampini } 107808122e43SStefano Zampini 107908122e43SStefano Zampini if (mss) { /* multilevel */ 108008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 108108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 108208122e43SStefano Zampini } 108308122e43SStefano Zampini 1084ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 108508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 108608122e43SStefano Zampini const PetscInt *idxs; 1087f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1088862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 108908122e43SStefano Zampini PetscBLASInt B_N; 1090aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 109108122e43SStefano Zampini 1092862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1093ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1094f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1095f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10969ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1097aff50787SStefano Zampini PetscInt j,k; 1098aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1099aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1100aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110108122e43SStefano Zampini } 110208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1103aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1104aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1105aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1106aff50787SStefano Zampini } 110708122e43SStefano Zampini } 110808122e43SStefano Zampini } else { 110908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111108122e43SStefano Zampini } 11128bec7fa6SStefano Zampini } else { 1113f6f667cfSStefano Zampini S = Sarray + cumarray; 1114f6f667cfSStefano Zampini St = Starray + cumarray; 11158bec7fa6SStefano Zampini } 1116aff50787SStefano Zampini /* see if we can save some work */ 1117aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1118aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1119aff50787SStefano Zampini } 1120aff50787SStefano Zampini 1121aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1122aff50787SStefano Zampini B_neigs = 0; 1123aff50787SStefano Zampini } else { 11249ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 112508122e43SStefano Zampini PetscBLASInt B_itype = 1; 1126f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11274c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11289552c7c7SStefano Zampini PetscInt nmin_s; 112908122e43SStefano Zampini 1130fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11318bec7fa6SStefano 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]]); 1132fd14bc51SStefano Zampini } 1133d16cbb6bSStefano Zampini 113408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1135d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1136d16cbb6bSStefano Zampini 1137d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 113808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1139f6f667cfSStefano 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)); 114008122e43SStefano Zampini #else 1141f6f667cfSStefano 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)); 114208122e43SStefano Zampini #endif 1143d16cbb6bSStefano Zampini } else { 1144d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1145d16cbb6bSStefano Zampini B_IL = 1; 1146d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1147d16cbb6bSStefano 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)); 1148d16cbb6bSStefano Zampini #else 1149d16cbb6bSStefano 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)); 1150d16cbb6bSStefano Zampini #endif 1151d16cbb6bSStefano Zampini } 115208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 115308122e43SStefano Zampini if (B_ierr) { 115408122e43SStefano Zampini if (B_ierr < 0 ) { 115508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 115608122e43SStefano Zampini } else if (B_ierr <= B_N) { 115708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 115808122e43SStefano Zampini } else { 11599552c7c7SStefano 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); 116008122e43SStefano Zampini } 116108122e43SStefano Zampini } 116208122e43SStefano Zampini 116308122e43SStefano Zampini if (B_neigs > nmax) { 1164fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1165fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1166fd14bc51SStefano Zampini } 1167f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 116808122e43SStefano Zampini B_neigs = nmax; 116908122e43SStefano Zampini } 117008122e43SStefano Zampini 11719552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11729552c7c7SStefano Zampini if (B_neigs < nmin_s) { 117308122e43SStefano Zampini PetscBLASInt B_neigs2; 117408122e43SStefano Zampini 1175f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1176f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1177fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1178fd14bc51SStefano 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); 1179fd14bc51SStefano Zampini } 11809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 118108122e43SStefano Zampini PetscInt j; 118208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118408122e43SStefano Zampini } 118508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118708122e43SStefano Zampini } 118808122e43SStefano Zampini } else { 118908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119108122e43SStefano Zampini } 119208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 119308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1194f6f667cfSStefano 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)); 119508122e43SStefano Zampini #else 1196f6f667cfSStefano 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)); 119708122e43SStefano Zampini #endif 119808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 119908122e43SStefano Zampini B_neigs += B_neigs2; 120008122e43SStefano Zampini } 120108122e43SStefano Zampini if (B_ierr) { 120208122e43SStefano Zampini if (B_ierr < 0 ) { 120308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 120408122e43SStefano Zampini } else if (B_ierr <= B_N) { 120508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 120608122e43SStefano Zampini } else { 12079552c7c7SStefano 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); 120808122e43SStefano Zampini } 120908122e43SStefano Zampini } 1210fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1211ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 121208122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 121308122e43SStefano Zampini if (eigs[j] == 0.0) { 1214ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 121508122e43SStefano Zampini } else { 1216ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1217fd14bc51SStefano Zampini } 121808122e43SStefano Zampini } 121908122e43SStefano Zampini } 122008122e43SStefano Zampini } else { 122108122e43SStefano Zampini /* TODO */ 122208122e43SStefano Zampini } 1223aff50787SStefano Zampini } 12246c3e6151SStefano Zampini /* change the basis back to the original one */ 12256c3e6151SStefano Zampini if (sub_schurs->change) { 122672b8c272SStefano Zampini Mat change,phi,phit; 12276c3e6151SStefano Zampini 12286c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12296c3e6151SStefano Zampini PetscInt ii; 12306c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12316c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12326c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12336c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12346c3e6151SStefano Zampini } 12356c3e6151SStefano Zampini } 12366c3e6151SStefano Zampini } 123772b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12386c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 123972b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12406c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12416c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12426c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12436c3e6151SStefano Zampini } 12448bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12458bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12469162d606SStefano Zampini if (B_neigs) { 12479162d606SStefano 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); 1248fd14bc51SStefano Zampini 1249fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12509552c7c7SStefano Zampini PetscInt ii; 12519552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1252ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12539552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1254ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1255ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1256ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1257ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1258ac47001eSStefano Zampini #else 1259ac47001eSStefano 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); 1260ac47001eSStefano Zampini #endif 12619552c7c7SStefano Zampini } 12629552c7c7SStefano Zampini } 1263fd14bc51SStefano Zampini } 12649162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12659162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12669162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12679162d606SStefano Zampini cum++; 126808122e43SStefano Zampini } 126908122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 127008122e43SStefano Zampini /* shift for next computation */ 127108122e43SStefano Zampini cumarray += subset_size*subset_size; 127208122e43SStefano Zampini } 1273fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1274fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1275fd14bc51SStefano Zampini } 127608122e43SStefano Zampini 127708122e43SStefano Zampini if (mss) { 127808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 127908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1280f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1281f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1282f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 128308122e43SStefano Zampini } 1284f6f667cfSStefano Zampini if (allocated_S_St) { 1285f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1286f6f667cfSStefano Zampini } 1287f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 128808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 128908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 129008122e43SStefano Zampini #endif 129108122e43SStefano Zampini if (pcbddc->dbg_flag) { 12921b968477SStefano Zampini PetscInt maxneigs_r; 129308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12949b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 129508122e43SStefano Zampini } 129608122e43SStefano Zampini PetscFunctionReturn(0); 129708122e43SStefano Zampini } 1298b1b3d7a2SStefano Zampini 1299674ae819SStefano Zampini #undef __FUNCT__ 1300c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1301c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1302c8587f34SStefano Zampini { 1303c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13048629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1305c8587f34SStefano Zampini PetscErrorCode ierr; 1306c8587f34SStefano Zampini 1307c8587f34SStefano Zampini PetscFunctionBegin; 1308f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 13095e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1310c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1311c8587f34SStefano Zampini 1312684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 13130fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1314684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1315c8587f34SStefano Zampini 1316c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1317b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1318c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1319c8587f34SStefano Zampini } 1320c8587f34SStefano Zampini 13218629588bSStefano Zampini /* 13228629588bSStefano Zampini Setup local correction and local part of coarse basis. 13238629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13248629588bSStefano Zampini */ 132547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13268629588bSStefano Zampini 13278629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13288629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13298629588bSStefano Zampini 13308629588bSStefano Zampini /* free */ 13318629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1332c8587f34SStefano Zampini PetscFunctionReturn(0); 1333c8587f34SStefano Zampini } 1334c8587f34SStefano Zampini 1335c8587f34SStefano Zampini #undef __FUNCT__ 1336674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1337674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1338674ae819SStefano Zampini { 1339674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1340674ae819SStefano Zampini PetscErrorCode ierr; 1341674ae819SStefano Zampini 1342674ae819SStefano Zampini PetscFunctionBegin; 1343674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1344674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 134530368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1346674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1347674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1348785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1349674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1350f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1351f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1352785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 135363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 135463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1355674ae819SStefano Zampini PetscFunctionReturn(0); 1356674ae819SStefano Zampini } 1357674ae819SStefano Zampini 1358674ae819SStefano Zampini #undef __FUNCT__ 1359674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1360674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1361674ae819SStefano Zampini { 1362674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13634f1b2e48SStefano Zampini PetscInt i; 1364674ae819SStefano Zampini PetscErrorCode ierr; 1365674ae819SStefano Zampini 1366674ae819SStefano Zampini PetscFunctionBegin; 1367b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1368674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1369674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1370674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13714f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13724f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13734f1b2e48SStefano Zampini } 13744f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1375b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1376674ae819SStefano Zampini PetscFunctionReturn(0); 1377674ae819SStefano Zampini } 1378674ae819SStefano Zampini 1379674ae819SStefano Zampini #undef __FUNCT__ 1380674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1381674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1382674ae819SStefano Zampini { 1383674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1384674ae819SStefano Zampini PetscErrorCode ierr; 1385674ae819SStefano Zampini 1386674ae819SStefano Zampini PetscFunctionBegin; 1387674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 138858da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1389ca92afb2SStefano Zampini PetscScalar *array; 139006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 139106656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 139258da7f69SStefano Zampini } 1393674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1394674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 139515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 139615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1397674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1398674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1399674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 140006656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1401674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1402674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14038ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1404674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1405674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1406674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1407f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1408f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1409f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1410f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1411727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 14120e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1413f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 141470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 14156e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 141681d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14170369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14188b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 14194f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14208b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1421ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1422ca92afb2SStefano Zampini PetscInt i; 1423ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1424ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1425ca92afb2SStefano Zampini } 1426ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1427ca92afb2SStefano Zampini } 14284f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1429674ae819SStefano Zampini PetscFunctionReturn(0); 1430674ae819SStefano Zampini } 1431674ae819SStefano Zampini 1432674ae819SStefano Zampini #undef __FUNCT__ 1433f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1434f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14356bfb1811SStefano Zampini { 14366bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14376bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14386bfb1811SStefano Zampini VecType impVecType; 14394f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14406bfb1811SStefano Zampini PetscErrorCode ierr; 14416bfb1811SStefano Zampini 14426bfb1811SStefano Zampini PetscFunctionBegin; 1443f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1444019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1445f4ddd8eeSStefano Zampini } 1446e7b262bdSStefano Zampini /* get sizes */ 14474f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1448b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14496bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1450e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1451e7b262bdSStefano Zampini /* R nodes */ 1452e7b262bdSStefano Zampini old_size = -1; 1453e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1454e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1455e7b262bdSStefano Zampini } 1456e7b262bdSStefano Zampini if (n_R != old_size) { 1457e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1458e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14596bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14606bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14616bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14626bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1463e7b262bdSStefano Zampini } 1464e7b262bdSStefano Zampini /* local primal dofs */ 1465e7b262bdSStefano Zampini old_size = -1; 1466e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1467e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1468e7b262bdSStefano Zampini } 1469e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1470e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 147183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1472e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14736bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1474e7b262bdSStefano Zampini } 1475e7b262bdSStefano Zampini /* local explicit constraints */ 1476e7b262bdSStefano Zampini old_size = -1; 1477e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1478e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1479e7b262bdSStefano Zampini } 1480e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1481e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 148283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 148383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 148483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 148583b7ccabSStefano Zampini } 14866bfb1811SStefano Zampini PetscFunctionReturn(0); 14876bfb1811SStefano Zampini } 14886bfb1811SStefano Zampini 14896bfb1811SStefano Zampini #undef __FUNCT__ 149047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 149147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 149288ebb749SStefano Zampini { 149325084f0cSStefano Zampini PetscErrorCode ierr; 149425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 149588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 149688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1497d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 149825084f0cSStefano Zampini /* submatrices of local problem */ 149980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 150006656605SStefano Zampini /* submatrices of local coarse problem */ 150106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 150225084f0cSStefano Zampini /* working matrices */ 150306656605SStefano Zampini Mat C_CR; 150425084f0cSStefano Zampini /* additional working stuff */ 150506656605SStefano Zampini PC pc_R; 15064f1b2e48SStefano Zampini Mat F; 1507*5cbda25cSStefano Zampini Vec dummy_vec; 1508a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 150925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 151006656605SStefano Zampini PetscScalar *work; 151106656605SStefano Zampini PetscInt *idx_V_B; 1512ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 151306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1514ffd830a3SStefano Zampini 151525084f0cSStefano Zampini /* some shortcuts to scalars */ 151606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 151788ebb749SStefano Zampini 151888ebb749SStefano Zampini PetscFunctionBegin; 1519ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1520ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1521ffd830a3SStefano Zampini } 1522ffd830a3SStefano Zampini 1523ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1524b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15254f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1526b371cd4fSStefano Zampini n_B = pcis->n_B; 1527b371cd4fSStefano Zampini n_D = pcis->n - n_B; 152888ebb749SStefano Zampini n_R = pcis->n - n_vertices; 152988ebb749SStefano Zampini 153088ebb749SStefano Zampini /* vertices in boundary numbering */ 1531785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15320e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 153388ebb749SStefano Zampini if (i != n_vertices) { 153422d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 153588ebb749SStefano Zampini } 153688ebb749SStefano Zampini 153706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1538019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 153906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 154006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 154106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 154206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 154306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 154406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 154506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 154606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 154706656605SStefano Zampini 154806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 154906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 155106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 155206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1553ffd830a3SStefano Zampini lda_rhs = n_R; 1554a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 155506656605SStefano Zampini if (isLU || isILU || isCHOL) { 155606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1557df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1558df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1559d62866d3SStefano Zampini MatFactorType type; 1560d62866d3SStefano Zampini 1561df4d28bfSStefano Zampini F = reuse_solver->F; 15626816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1563d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1564ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1565df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 156606656605SStefano Zampini } else { 156706656605SStefano Zampini F = NULL; 156806656605SStefano Zampini } 156906656605SStefano Zampini 1570ffd830a3SStefano Zampini /* allocate workspace */ 1571ffd830a3SStefano Zampini n = 0; 1572ffd830a3SStefano Zampini if (n_constraints) { 1573ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1574ffd830a3SStefano Zampini } 1575ffd830a3SStefano Zampini if (n_vertices) { 1576ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1577ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1578ffd830a3SStefano Zampini } 1579ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1580ffd830a3SStefano Zampini 1581*5cbda25cSStefano Zampini /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */ 1582*5cbda25cSStefano Zampini dummy_vec = NULL; 1583*5cbda25cSStefano Zampini if (need_benign_correction && lda_rhs != n_R && F) { 1584*5cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,lda_rhs,work,&dummy_vec);CHKERRQ(ierr); 1585*5cbda25cSStefano Zampini } 1586*5cbda25cSStefano Zampini 158788ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 158888ebb749SStefano Zampini if (n_constraints) { 158972b8c272SStefano Zampini Mat M1,M2,M3,C_B; 159006656605SStefano Zampini IS is_aux; 159180677318SStefano Zampini PetscScalar *array,*array2; 159206656605SStefano Zampini 1593f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 159480677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 159588ebb749SStefano Zampini 159625084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 159725084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15988ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 159972b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 160088ebb749SStefano Zampini 160180677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 160280677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1603ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 160488ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 160506656605SStefano Zampini const PetscScalar *row_cmat_values; 160606656605SStefano Zampini const PetscInt *row_cmat_indices; 160706656605SStefano Zampini PetscInt size_of_constraint,j; 160888ebb749SStefano Zampini 160906656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161006656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1611ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 161206656605SStefano Zampini } 161306656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 161406656605SStefano Zampini } 1615ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 161606656605SStefano Zampini if (F) { 161706656605SStefano Zampini Mat B; 161806656605SStefano Zampini 1619ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1620a3df083aSStefano Zampini if (need_benign_correction) { 1621df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1622a3df083aSStefano Zampini 162372b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 162472b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1625a3df083aSStefano Zampini } 162680677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1627a3df083aSStefano Zampini if (need_benign_correction) { 1628a3df083aSStefano Zampini PetscScalar *marr; 1629df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1630a3df083aSStefano Zampini 1631a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1632*5cbda25cSStefano Zampini if (lda_rhs != n_R) { 1633*5cbda25cSStefano Zampini for (i=0;i<n_constraints;i++) { 1634*5cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 1635*5cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1636*5cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 1637*5cbda25cSStefano Zampini } 1638*5cbda25cSStefano Zampini } else { 1639a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1640a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1641*5cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1642a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1643a3df083aSStefano Zampini } 1644*5cbda25cSStefano Zampini } 1645a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1646a3df083aSStefano Zampini } 164706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 164806656605SStefano Zampini } else { 164980677318SStefano Zampini PetscScalar *marr; 165080677318SStefano Zampini 165180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 165206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1653ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1654ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 165506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 165606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 165706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 165806656605SStefano Zampini } 165980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 166006656605SStefano Zampini } 166180677318SStefano Zampini if (!pcbddc->switch_static) { 166280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 166380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 166480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 166580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1666ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 166780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 166880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 166980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 167180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 167280677318SStefano Zampini } 167380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 167480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 167572b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 167680677318SStefano Zampini } else { 1677ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1678ffd830a3SStefano Zampini IS dummy; 1679ffd830a3SStefano Zampini 1680ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 168172b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1682ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1683ffd830a3SStefano Zampini } else { 168480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 168580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1686ffd830a3SStefano Zampini } 168725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 168880677318SStefano Zampini } 168980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 169080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 169180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 169206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 169306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 169480677318SStefano Zampini if (isCHOL) { 169580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 169680677318SStefano Zampini } else { 169725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 169880677318SStefano Zampini } 169980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 170006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 170125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 170225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 170325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 170480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 170572b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 170672b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 170706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 170806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1709f4ddd8eeSStefano Zampini } 1710fc227af8SStefano Zampini 1711fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 171288ebb749SStefano Zampini if (n_vertices) { 171306656605SStefano Zampini IS is_aux; 17143a50541eSStefano Zampini 1715df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 17166816873aSStefano Zampini IS tis; 17176816873aSStefano Zampini 17186816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17196816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17206816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17216816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17226816873aSStefano Zampini } else { 17233a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17246816873aSStefano Zampini } 17259577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17269577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 172704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 172825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 172988ebb749SStefano Zampini } 173088ebb749SStefano Zampini 173188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1732f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 173306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 173406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 173506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 173606656605SStefano Zampini } 1737f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 173806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 173906656605SStefano Zampini PetscScalar *marray; 174006656605SStefano Zampini 174106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 174206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1743f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1744f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1745f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1746f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1747f4ddd8eeSStefano Zampini } 1748f4ddd8eeSStefano Zampini } 174906656605SStefano Zampini 1750f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 175106656605SStefano Zampini PetscScalar *marray; 175288ebb749SStefano Zampini 175306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17548eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 175506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 175688ebb749SStefano Zampini } 17573301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 175806656605SStefano Zampini n *= 2; 175988ebb749SStefano Zampini } 176006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 176106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 176206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17638eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 176406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 176506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 176688ebb749SStefano Zampini } 17673301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 176806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17698eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 177006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 177106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 177288ebb749SStefano Zampini } 177388ebb749SStefano Zampini } else { 1774c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1775c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17761b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1777c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1778c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1779c0553b1fSStefano Zampini } 178088ebb749SStefano Zampini } 178106656605SStefano Zampini } 1782019a44ceSStefano Zampini 178306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17844f1b2e48SStefano Zampini p0_lidx_I = NULL; 17854f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1786d12edf2fSStefano Zampini const PetscInt *idxs; 1787d12edf2fSStefano Zampini 1788d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17894f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17904f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17914f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17924f1b2e48SStefano Zampini } 1793d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1794d12edf2fSStefano Zampini } 1795d16cbb6bSStefano Zampini 179606656605SStefano Zampini /* vertices */ 179706656605SStefano Zampini if (n_vertices) { 179816f15bc4SStefano Zampini 1799ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 180004708bb6SStefano Zampini 180116f15bc4SStefano Zampini if (n_R) { 180214393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 180306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 180416f15bc4SStefano Zampini PetscScalar *x,*y; 180504708bb6SStefano Zampini PetscBool isseqaij; 180606656605SStefano Zampini 180721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 180814393ed6SStefano Zampini if (need_benign_correction) { 180914393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 181014393ed6SStefano Zampini IS is_p0; 181114393ed6SStefano Zampini PetscInt *idxs_p0,n; 181214393ed6SStefano Zampini 181314393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 181414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 181514393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 181614393ed6SStefano Zampini if (n != pcbddc->benign_n) { 181714393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 181814393ed6SStefano Zampini } 181914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 182014393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 182114393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 182214393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 182314393ed6SStefano Zampini } 182414393ed6SStefano Zampini 1825ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1826ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1827ffd830a3SStefano Zampini } else { 1828ca92afb2SStefano Zampini PetscScalar *av,*array; 1829ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1830ca92afb2SStefano Zampini PetscInt n; 1831ca92afb2SStefano Zampini PetscBool flg_row; 1832ffd830a3SStefano Zampini 1833ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1834ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1835ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1836ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1837ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1838ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1839ca92afb2SStefano Zampini PetscInt j; 1840ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1841ffd830a3SStefano Zampini } 1842ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1843ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1844ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1845ffd830a3SStefano Zampini } 1846ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1847a3df083aSStefano Zampini if (need_benign_correction) { 1848df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1849a3df083aSStefano Zampini PetscScalar *marr; 1850a3df083aSStefano Zampini 1851a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 185214393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 185314393ed6SStefano Zampini 185414393ed6SStefano Zampini | 0 0 0 | (V) 185514393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 185614393ed6SStefano Zampini | 0 0 -1 | (p0) 185714393ed6SStefano Zampini 185814393ed6SStefano Zampini */ 1859df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 186014393ed6SStefano Zampini const PetscScalar *vals; 186114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 186214393ed6SStefano Zampini PetscInt n,j,nz; 186314393ed6SStefano Zampini 1864df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1865df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 186714393ed6SStefano Zampini for (j=0;j<n;j++) { 186814393ed6SStefano Zampini PetscScalar val = vals[j]; 186914393ed6SStefano Zampini PetscInt k,col = idxs[j]; 187014393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 187114393ed6SStefano Zampini } 187214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1873df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 187414393ed6SStefano Zampini } 187572b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 187672b8c272SStefano Zampini } 187772b8c272SStefano Zampini if (F) { 187814393ed6SStefano Zampini /* need to correct the rhs */ 187972b8c272SStefano Zampini if (need_benign_correction) { 188072b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 188172b8c272SStefano Zampini PetscScalar *marr; 188272b8c272SStefano Zampini 188372b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 1884*5cbda25cSStefano Zampini if (lda_rhs != n_R) { 1885*5cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 1886*5cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 1887*5cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1888*5cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 1889*5cbda25cSStefano Zampini } 1890*5cbda25cSStefano Zampini } else { 1891a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1892a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1893*5cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1894a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1895a3df083aSStefano Zampini } 1896*5cbda25cSStefano Zampini } 1897a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1898a3df083aSStefano Zampini } 189906656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 190014393ed6SStefano Zampini /* need to correct the solution */ 1901a3df083aSStefano Zampini if (need_benign_correction) { 1902df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1903a3df083aSStefano Zampini PetscScalar *marr; 1904a3df083aSStefano Zampini 1905a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1906*5cbda25cSStefano Zampini if (lda_rhs != n_R) { 1907*5cbda25cSStefano Zampini for (i=0;i<n_vertices;i++) { 1908*5cbda25cSStefano Zampini ierr = VecPlaceArray(dummy_vec,marr+i*lda_rhs);CHKERRQ(ierr); 1909*5cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1910*5cbda25cSStefano Zampini ierr = VecResetArray(dummy_vec);CHKERRQ(ierr); 1911*5cbda25cSStefano Zampini } 1912*5cbda25cSStefano Zampini } else { 1913a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1914a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1915*5cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 1916a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1917a3df083aSStefano Zampini } 1918*5cbda25cSStefano Zampini } 1919a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1920a3df083aSStefano Zampini } 192106656605SStefano Zampini } else { 192206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 192306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1924ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1925ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 192606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 192706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 192806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 192906656605SStefano Zampini } 193006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 193106656605SStefano Zampini } 193280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1933ffd830a3SStefano Zampini /* S_VV and S_CV */ 193406656605SStefano Zampini if (n_constraints) { 193506656605SStefano Zampini Mat B; 193680677318SStefano Zampini 1937ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 193880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1939ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1940ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 194180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 194480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 194580677318SStefano Zampini } 1946ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 194780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 194880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1949ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 195080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 195106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1952ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1953ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 195406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195506656605SStefano Zampini } 195604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 195704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 195804708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 195904708bb6SStefano Zampini } 1960ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1961ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1962ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1963ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1964ffd830a3SStefano Zampini } 196506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 196614393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 196714393ed6SStefano Zampini if (need_benign_correction) { 1968df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 196914393ed6SStefano Zampini PetscScalar *marr,*sums; 197014393ed6SStefano Zampini 197114393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 197214393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1973df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 197414393ed6SStefano Zampini const PetscScalar *vals; 197514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 197614393ed6SStefano Zampini PetscInt n,j,nz; 197714393ed6SStefano Zampini 1978df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1979df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 198014393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 198114393ed6SStefano Zampini PetscInt k; 198214393ed6SStefano Zampini sums[j] = 0.; 198314393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 198414393ed6SStefano Zampini } 198514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 198614393ed6SStefano Zampini for (j=0;j<n;j++) { 198714393ed6SStefano Zampini PetscScalar val = vals[j]; 198814393ed6SStefano Zampini PetscInt k; 198914393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 199014393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 199114393ed6SStefano Zampini } 199214393ed6SStefano Zampini } 199314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1994df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 199514393ed6SStefano Zampini } 199614393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 199714393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 199814393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 199914393ed6SStefano Zampini } 200080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 200206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 200306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 200406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 200506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 200606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 200706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2008d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 2009019a44ceSStefano Zampini } else { 2010d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2011d16cbb6bSStefano Zampini } 201221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 2013d16cbb6bSStefano Zampini 201406656605SStefano Zampini /* coarse basis functions */ 201506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 201616f15bc4SStefano Zampini PetscScalar *y; 201716f15bc4SStefano Zampini 2018ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 201906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 202106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 202306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 202406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 202506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 202606656605SStefano Zampini 202706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20284f1b2e48SStefano Zampini PetscInt j; 20294f1b2e48SStefano Zampini 203006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 203206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 203406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20354f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 203606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 203706656605SStefano Zampini } 203806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 203906656605SStefano Zampini } 204004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 204104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 204206656605SStefano Zampini } 2043*5cbda25cSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 204406656605SStefano Zampini 204506656605SStefano Zampini if (n_constraints) { 204606656605SStefano Zampini Mat B; 204706656605SStefano Zampini 2048ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 204906656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 205106656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 205206656605SStefano Zampini if (n_vertices) { 205380677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 205480677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 205580677318SStefano Zampini } else { 205680677318SStefano Zampini Mat S_VCt; 205780677318SStefano Zampini 2058ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2059ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206072b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2061ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2062ffd830a3SStefano Zampini } 206380677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 206480677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 206580677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 206680677318SStefano Zampini } 206706656605SStefano Zampini } 206806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 206906656605SStefano Zampini /* coarse basis functions */ 207006656605SStefano Zampini for (i=0;i<n_constraints;i++) { 207106656605SStefano Zampini PetscScalar *y; 207206656605SStefano Zampini 2073ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 207406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 207606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 207806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 207906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 208006656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20814f1b2e48SStefano Zampini PetscInt j; 20824f1b2e48SStefano Zampini 208306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 208406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 208506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20884f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 208906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 209006656605SStefano Zampini } 209106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209206656605SStefano Zampini } 209306656605SStefano Zampini } 209480677318SStefano Zampini if (n_constraints) { 209580677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 209680677318SStefano Zampini } 20974f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 209872b8c272SStefano Zampini 209972b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 210072b8c272SStefano Zampini if (pcbddc->benign_n) { 210172b8c272SStefano Zampini Mat B0_B,B0_BPHI; 210272b8c272SStefano Zampini IS is_dummy; 210372b8c272SStefano Zampini PetscScalar *data; 210472b8c272SStefano Zampini PetscInt j; 210572b8c272SStefano Zampini 210672b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 210772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 210872b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 210972b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 211072b8c272SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 211172b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 211272b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 211372b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 211472b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 211572b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 211672b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 211772b8c272SStefano Zampini } 211872b8c272SStefano Zampini } 211972b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 212072b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 212172b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 212272b8c272SStefano Zampini } 2123019a44ceSStefano Zampini 212406656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 21253301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2126ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2127ffd830a3SStefano Zampini PetscScalar *marray; 212806656605SStefano Zampini 212906656605SStefano Zampini if (n_constraints) { 2130ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 213106656605SStefano Zampini 2132ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 213306656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2134ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 213516f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 213606656605SStefano Zampini if (n_vertices) { 2137ffd830a3SStefano Zampini Mat S_VCT; 213806656605SStefano Zampini 213906656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2140ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 214116f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 214206656605SStefano Zampini } 2143ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 214406656605SStefano Zampini } 214516f15bc4SStefano Zampini if (n_vertices && n_R) { 2146ffd830a3SStefano Zampini PetscScalar *av,*marray; 2147ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2148ffd830a3SStefano Zampini PetscInt n; 2149ffd830a3SStefano Zampini PetscBool flg_row; 215006656605SStefano Zampini 2151ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2152ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2153ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2154ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2155ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2156ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2157ffd830a3SStefano Zampini PetscInt j; 2158ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2159ffd830a3SStefano Zampini } 2160ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2161ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2162ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 216306656605SStefano Zampini } 216406656605SStefano Zampini 2165ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2166ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2167ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2168ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2169ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 217006656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 217106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 217306656605SStefano Zampini } 2174ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2175ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2176ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2177ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2178ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2179ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2180ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2181ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 218206656605SStefano Zampini } 2183ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 218406656605SStefano Zampini /* coarse basis functions */ 218506656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 218606656605SStefano Zampini PetscScalar *y; 218706656605SStefano Zampini 2188ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 218906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 219106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 219306656605SStefano Zampini if (i<n_vertices) { 219406656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 219506656605SStefano Zampini } 219606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 219706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 219806656605SStefano Zampini 219906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 220006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 220206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 220406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 220506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 220606656605SStefano Zampini } 220706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 220806656605SStefano Zampini } 2209ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2210ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 221106656605SStefano Zampini } 2212d62866d3SStefano Zampini /* free memory */ 221388ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 221406656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 221506656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 221606656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 221706656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2218d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2219d62866d3SStefano Zampini if (n_vertices) { 2220d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2221d62866d3SStefano Zampini } 2222d62866d3SStefano Zampini if (n_constraints) { 2223d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2224d62866d3SStefano Zampini } 222588ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 222688ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 222788ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2228d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 222988ebb749SStefano Zampini Mat coarse_sub_mat; 223025084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 223188ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 223288ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 223388ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22348bec7fa6SStefano Zampini Mat C_B,CPHI; 22358bec7fa6SStefano Zampini IS is_dummy; 22368bec7fa6SStefano Zampini Vec mones; 223788ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 223888ebb749SStefano Zampini PetscReal real_value; 223988ebb749SStefano Zampini 2240a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2241a3df083aSStefano Zampini Mat A; 2242a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2243a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2244a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2245a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2246a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2247a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2248a3df083aSStefano Zampini } else { 224988ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 225088ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 225188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 225288ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2253a3df083aSStefano Zampini } 225488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 225588ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2256ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 225788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 225888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 225988ebb749SStefano Zampini } 226088ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 226188ebb749SStefano Zampini 226225084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22633301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 226425084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2265ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 226688ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 226788ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 226888ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 226988ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 227188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227288ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 227688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 227788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 227888ebb749SStefano Zampini } else { 227988ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 228088ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 228188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 228388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 228588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 228688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 228788ebb749SStefano Zampini } 228888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 228988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229088ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 229188ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 22924f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2293fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2294d12edf2fSStefano Zampini PetscScalar *data,*data2; 22954f1b2e48SStefano Zampini PetscInt j; 2296d12edf2fSStefano Zampini 22974f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2298fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2299d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2300d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2301d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2302d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 23034f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 23044f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2305d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 23064f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 23074f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 23084f1b2e48SStefano Zampini } 2309d12edf2fSStefano Zampini } 2310d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2311d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2312d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2313d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2314d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2315d12edf2fSStefano Zampini } 2316d12edf2fSStefano Zampini #if 0 2317d12edf2fSStefano Zampini { 2318d12edf2fSStefano Zampini PetscViewer viewer; 2319d12edf2fSStefano Zampini char filename[256]; 2320ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2321d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2322d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2323ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2324ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2325ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2326d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 232772b8c272SStefano Zampini if (save_change) { 232872b8c272SStefano Zampini Mat phi_B; 232972b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 233072b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 233172b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 233272b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 233372b8c272SStefano Zampini } else { 2334ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2335ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 233672b8c272SStefano Zampini } 2337ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2338ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2339ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2340ffd830a3SStefano Zampini } 2341ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2342ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2343ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2344ffd830a3SStefano Zampini } 234572b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2346ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2347ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2348ffd830a3SStefano Zampini } 2349d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2350d12edf2fSStefano Zampini } 2351d12edf2fSStefano Zampini #endif 235281d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23538bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23541575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 235506656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23568bec7fa6SStefano Zampini 23578bec7fa6SStefano Zampini /* check constraints */ 23584f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23598bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 23608bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 23618bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 23628bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23638bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23648bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23658bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2366bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2367ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2368bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2369bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2370bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2371bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2372bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 237388ebb749SStefano Zampini } 23748bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23758bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23768bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23778bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2378d12edf2fSStefano Zampini } 237925084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 238088ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 238188ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 238288ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 238388ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 238488ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 238588ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 238688ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 238788ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 238888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 238988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2390ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 239188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 239288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 239388ebb749SStefano Zampini } 239488ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 239588ebb749SStefano Zampini } 23968629588bSStefano Zampini /* get back data */ 23978629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 239888ebb749SStefano Zampini PetscFunctionReturn(0); 239988ebb749SStefano Zampini } 240088ebb749SStefano Zampini 240188ebb749SStefano Zampini #undef __FUNCT__ 2402d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2403d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2404aa0d41d4SStefano Zampini { 2405d65f70fdSStefano Zampini Mat *work_mat; 2406d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2407d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2408d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2409aa0d41d4SStefano Zampini PetscErrorCode ierr; 2410aa0d41d4SStefano Zampini 2411aa0d41d4SStefano Zampini PetscFunctionBegin; 2412d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2413d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2414d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2415d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2416aa0d41d4SStefano Zampini 2417d65f70fdSStefano Zampini if (!rsorted) { 2418906d46d4SStefano Zampini const PetscInt *idxs; 2419906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2420aa0d41d4SStefano Zampini 2421d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2422d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2423d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2424d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2425aa0d41d4SStefano Zampini } 2426d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2427d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2428d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2429d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2430aa0d41d4SStefano Zampini } 2431d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2432d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2433d65f70fdSStefano Zampini } else { 2434d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2435d65f70fdSStefano Zampini isrow_s = isrow; 2436aa0d41d4SStefano Zampini } 2437906d46d4SStefano Zampini 2438d65f70fdSStefano Zampini if (!csorted) { 2439d65f70fdSStefano Zampini if (isrow == iscol) { 2440d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2441d65f70fdSStefano Zampini iscol_s = isrow_s; 2442d65f70fdSStefano Zampini } else { 2443d65f70fdSStefano Zampini const PetscInt *idxs; 2444d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2445906d46d4SStefano Zampini 2446d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2447d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2448d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2449d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2450d65f70fdSStefano Zampini } 2451d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2452d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2453d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2454d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2455d65f70fdSStefano Zampini } 2456d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2457d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2458d65f70fdSStefano Zampini } 2459d65f70fdSStefano Zampini } else { 2460d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2461d65f70fdSStefano Zampini iscol_s = iscol; 2462d65f70fdSStefano Zampini } 2463d65f70fdSStefano Zampini 2464d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2465d65f70fdSStefano Zampini 2466d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2467906d46d4SStefano Zampini Mat new_mat; 2468d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2469906d46d4SStefano Zampini 2470d65f70fdSStefano Zampini if (!rsorted) { 2471d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2472d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2473d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2474d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2475906d46d4SStefano Zampini } 2476d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2477d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2478d65f70fdSStefano Zampini } else { 2479d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2480906d46d4SStefano Zampini } 2481d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2482d65f70fdSStefano Zampini 2483d65f70fdSStefano Zampini if (!csorted) { 2484d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2485d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2486d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2487d65f70fdSStefano Zampini } else { 2488d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2489d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2490d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2491d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2492d65f70fdSStefano Zampini } 2493d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2494d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2495d65f70fdSStefano Zampini } 2496d65f70fdSStefano Zampini } else { 2497d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2498d65f70fdSStefano Zampini } 2499d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2500d65f70fdSStefano Zampini 2501d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2502d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2503d65f70fdSStefano Zampini work_mat[0] = new_mat; 2504d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2505d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2506d65f70fdSStefano Zampini } 2507d65f70fdSStefano Zampini 2508d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2509d65f70fdSStefano Zampini *B = work_mat[0]; 2510d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2511d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2512d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2513d65f70fdSStefano Zampini PetscFunctionReturn(0); 2514d65f70fdSStefano Zampini } 2515d65f70fdSStefano Zampini 2516d65f70fdSStefano Zampini #undef __FUNCT__ 25175e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 25185e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2519aa0d41d4SStefano Zampini { 2520aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 25215e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2522d65f70fdSStefano Zampini Mat new_mat; 25235e8657edSStefano Zampini IS is_local,is_global; 2524d65f70fdSStefano Zampini PetscInt local_size; 2525d65f70fdSStefano Zampini PetscBool isseqaij; 2526aa0d41d4SStefano Zampini PetscErrorCode ierr; 2527aa0d41d4SStefano Zampini 2528aa0d41d4SStefano Zampini PetscFunctionBegin; 2529aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 25305e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25315e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2532b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2533aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2534d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2535aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2536906d46d4SStefano Zampini 2537906d46d4SStefano Zampini /* check */ 2538906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2539906d46d4SStefano Zampini Vec x,x_change; 2540906d46d4SStefano Zampini PetscReal error; 2541906d46d4SStefano Zampini 25425e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2543906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25445e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2545e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2546e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2547d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2548e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2549e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2550906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2551906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2552906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2553906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2554906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2555906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2556906d46d4SStefano Zampini } 2557906d46d4SStefano Zampini 255822d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25599b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 256022d5777bSStefano Zampini if (isseqaij) { 25611cf9b237SStefano Zampini Mat M; 25621cf9b237SStefano Zampini 25631cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 25641cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 25651cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2566aa0d41d4SStefano Zampini } else { 25671cf9b237SStefano Zampini Mat work_mat,M; 25681cf9b237SStefano Zampini 2569aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 25701cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 25711cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 25721cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2573aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2574aa0d41d4SStefano Zampini } 25753301b35fSStefano Zampini if (matis->A->symmetric_set) { 25763301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2577e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25783301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2579e496cd5dSStefano Zampini #endif 25803301b35fSStefano Zampini } 258145a1bb75SStefano Zampini /* 258245a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2583d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 258445a1bb75SStefano Zampini */ 2585d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2586aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2587aa0d41d4SStefano Zampini } 2588aa0d41d4SStefano Zampini 2589aa0d41d4SStefano Zampini #undef __FUNCT__ 2590a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25918ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2592a64d13efSStefano Zampini { 2593a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2594a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2595d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 259653892102SStefano Zampini PetscInt *idx_R_local=NULL; 25973a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25983a50541eSStefano Zampini PetscInt vbs,bs; 25996816873aSStefano Zampini PetscBT bitmask=NULL; 2600a64d13efSStefano Zampini PetscErrorCode ierr; 2601a64d13efSStefano Zampini 2602a64d13efSStefano Zampini PetscFunctionBegin; 2603b23d619eSStefano Zampini /* 2604b23d619eSStefano Zampini No need to setup local scatters if 2605b23d619eSStefano Zampini - primal space is unchanged 2606b23d619eSStefano Zampini AND 2607b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2608b23d619eSStefano Zampini AND 2609b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2610b23d619eSStefano Zampini */ 2611b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2612f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2613f4ddd8eeSStefano Zampini } 2614f4ddd8eeSStefano Zampini /* destroy old objects */ 2615f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2616f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2617f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2618a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2619b371cd4fSStefano Zampini n_B = pcis->n_B; 2620b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2621b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 26223a50541eSStefano Zampini 2623a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 26246816873aSStefano Zampini 262553892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2626df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2627854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2628a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2629a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 26300e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2631a64d13efSStefano Zampini } 2632a64d13efSStefano Zampini 2633a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26344641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26356816873aSStefano Zampini idx_R_local[n_R++] = i; 2636a64d13efSStefano Zampini } 2637a64d13efSStefano Zampini } 2638df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2639df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26406816873aSStefano Zampini 2641df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2642df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26436816873aSStefano Zampini } 26443a50541eSStefano Zampini 26453a50541eSStefano Zampini /* Block code */ 26463a50541eSStefano Zampini vbs = 1; 26473a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26483a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26493a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26503a50541eSStefano Zampini PetscInt *vary; 2651df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2652785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26533a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2654d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2655d3df7717SStefano 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 */ 26560e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2657d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26583a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26593a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26603a50541eSStefano Zampini break; 26613a50541eSStefano Zampini } 26623a50541eSStefano Zampini } 2663d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2664d3df7717SStefano Zampini } else { 2665d3df7717SStefano Zampini /* Verify directly the R set */ 2666d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2667d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2668d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2669d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2670d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2671d3df7717SStefano Zampini break; 2672d3df7717SStefano Zampini } 2673d3df7717SStefano Zampini } 2674d3df7717SStefano Zampini } 2675d3df7717SStefano Zampini } 26763a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26773a50541eSStefano Zampini vbs = bs; 26783a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26793a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26803a50541eSStefano Zampini } 26813a50541eSStefano Zampini } 26823a50541eSStefano Zampini } 26833a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2684df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2685df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 268653892102SStefano Zampini 2687df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2688df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 268953892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2690df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 269153892102SStefano Zampini } else { 26923a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 269353892102SStefano Zampini } 2694a64d13efSStefano Zampini 2695a64d13efSStefano Zampini /* print some info if requested */ 2696a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2697a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2698a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26991575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2700a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2701a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 27024f1b2e48SStefano 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); 2703a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2704a64d13efSStefano Zampini } 2705a64d13efSStefano Zampini 2706a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2707df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 27086816873aSStefano Zampini IS is_aux1,is_aux2; 27096816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 27106816873aSStefano Zampini 27113a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2712854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2713854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2714a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 27154641a718SStefano Zampini for (i=0; i<n_D; i++) { 27164641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 27174641a718SStefano Zampini } 2718a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2719a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27204641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 27214641a718SStefano Zampini aux_array1[j++] = i; 2722a64d13efSStefano Zampini } 2723a64d13efSStefano Zampini } 2724a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2725a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2726a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 27274641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 27284641a718SStefano Zampini aux_array2[j++] = i; 2729a64d13efSStefano Zampini } 2730a64d13efSStefano Zampini } 2731a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2732a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2733a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2734a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2735a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2736a64d13efSStefano Zampini 27378eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2738785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2739a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27404641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27414641a718SStefano Zampini aux_array1[j++] = i; 2742a64d13efSStefano Zampini } 2743a64d13efSStefano Zampini } 2744a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2745a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2746a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2747a64d13efSStefano Zampini } 27484641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27493a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2750d62866d3SStefano Zampini } else { 2751df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27526816873aSStefano Zampini IS tis; 27536816873aSStefano Zampini PetscInt schur_size; 27546816873aSStefano Zampini 2755df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27566816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2757df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27586816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27596816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27606816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27616816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27626816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2763d62866d3SStefano Zampini } 2764d62866d3SStefano Zampini } 2765a64d13efSStefano Zampini PetscFunctionReturn(0); 2766a64d13efSStefano Zampini } 2767a64d13efSStefano Zampini 2768304d26faSStefano Zampini 2769304d26faSStefano Zampini #undef __FUNCT__ 2770304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2771684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2772304d26faSStefano Zampini { 2773304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2774304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2775304d26faSStefano Zampini PC pc_temp; 2776304d26faSStefano Zampini Mat A_RR; 2777f4ddd8eeSStefano Zampini MatReuse reuse; 2778304d26faSStefano Zampini PetscScalar m_one = -1.0; 2779304d26faSStefano Zampini PetscReal value; 278004708bb6SStefano Zampini PetscInt n_D,n_R; 27819577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2782304d26faSStefano Zampini PetscErrorCode ierr; 2783e604994aSStefano Zampini /* prefixes stuff */ 2784312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2785e604994aSStefano Zampini size_t len; 2786304d26faSStefano Zampini 2787304d26faSStefano Zampini PetscFunctionBegin; 2788304d26faSStefano Zampini 2789e604994aSStefano Zampini /* compute prefixes */ 2790e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2791e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2792e604994aSStefano Zampini if (!pcbddc->current_level) { 2793e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2794e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2795e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2796e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2797e604994aSStefano Zampini } else { 2798e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2799312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2800e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2801e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2802312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2803312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 280434d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 280534d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2806e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2807e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2808e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2809e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2810e604994aSStefano Zampini } 2811e604994aSStefano Zampini 2812304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2813684f6988SStefano Zampini if (dirichlet) { 2814d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2815450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2816450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2817450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2818450f8f5eSStefano Zampini } 2819450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2820a3df083aSStefano Zampini Mat A_IIn; 2821a3df083aSStefano Zampini 2822a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2823a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2824a3df083aSStefano Zampini pcis->A_II = A_IIn; 2825a3df083aSStefano Zampini } 2826450f8f5eSStefano Zampini } 28273301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28283301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2829964fefecSStefano Zampini } 2830ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2831964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2832304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2833304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2834304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2835304d26faSStefano Zampini /* default */ 2836304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2837e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28389577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2839304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28409577ea80SStefano Zampini if (issbaij) { 28419577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28429577ea80SStefano Zampini } else { 2843304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28449577ea80SStefano Zampini } 2845304d26faSStefano Zampini /* Allow user's customization */ 2846304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2847304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2848304d26faSStefano Zampini } 2849d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2850df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2851df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2852d62866d3SStefano Zampini 2853df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2854d5574798SStefano Zampini } 2855304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2856304d26faSStefano Zampini if (!n_D) { 2857304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2858304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2859304d26faSStefano Zampini } 2860304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2861304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2862304d26faSStefano Zampini /* set ksp_D into pcis data */ 2863304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2864304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2865304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2866684f6988SStefano Zampini } 2867304d26faSStefano Zampini 2868304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2869684f6988SStefano Zampini A_RR = 0; 2870684f6988SStefano Zampini if (neumann) { 2871d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 287204708bb6SStefano Zampini PetscInt ibs,mbs; 287304708bb6SStefano Zampini PetscBool issbaij; 287404708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2875f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28768ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2877f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2878f4ddd8eeSStefano Zampini PetscInt nn_R; 287981d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2880f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2881f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2882f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2883f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2884f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2885f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2886f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2887727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2888f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2889f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2890f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2891f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2892f4ddd8eeSStefano Zampini } 2893f4ddd8eeSStefano Zampini } 2894f4ddd8eeSStefano Zampini /* last check */ 2895d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2896f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2897f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2898f4ddd8eeSStefano Zampini } 2899f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2900f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2901f4ddd8eeSStefano Zampini } 2902f4ddd8eeSStefano Zampini /* extract A_RR */ 2903af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2904af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 290504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 290604708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 290704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 290804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 290904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2910af732b37SStefano Zampini } else { 291104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 29126816873aSStefano Zampini } 291304708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 291404708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 291504708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 291604708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 291704708bb6SStefano Zampini } else { 291804708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 291904708bb6SStefano Zampini } 292004708bb6SStefano Zampini } 292116e386b8SStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 292216e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 292316e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 292416e386b8SStefano Zampini } else { 2925f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 292616e386b8SStefano Zampini } 29273301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 29283301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 29296816873aSStefano Zampini } 2930f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2931304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2932304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2933304d26faSStefano Zampini /* default */ 2934304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2935e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2936304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29379577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29389577ea80SStefano Zampini if (issbaij) { 29399577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29409577ea80SStefano Zampini } else { 2941304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29429577ea80SStefano Zampini } 2943304d26faSStefano Zampini /* Allow user's customization */ 2944304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2945304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2946304d26faSStefano Zampini } 2947304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2948304d26faSStefano Zampini if (!n_R) { 2949304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2950304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2951304d26faSStefano Zampini } 2952*5cbda25cSStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2953df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2954df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2955df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2956d62866d3SStefano Zampini 2957df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2958d62866d3SStefano Zampini } 2959304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2960304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2961684f6988SStefano Zampini } 2962304d26faSStefano Zampini 2963304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29640fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2965684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2966684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29671575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2968684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2969684f6988SStefano Zampini } 2970684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29710fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29720fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29730fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29740fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29750fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2976304d26faSStefano Zampini /* need to be adapted? */ 2977b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2978b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2979b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2980304d26faSStefano Zampini /* print info */ 2981304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2982e604994aSStefano 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); 2983304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2984304d26faSStefano Zampini } 2985b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2986298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2987304d26faSStefano Zampini } 2988684f6988SStefano Zampini } 2989684f6988SStefano Zampini if (neumann) { /* Neumann */ 29900fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 29910fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 29920fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 29930fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 29940fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2995304d26faSStefano Zampini /* need to be adapted? */ 2996b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2997b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2998304d26faSStefano Zampini /* print info */ 2999304d26faSStefano Zampini if (pcbddc->dbg_flag) { 3000e604994aSStefano 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); 3001304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 3002304d26faSStefano Zampini } 3003b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 3004298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 3005304d26faSStefano Zampini } 30060fccc4e9SStefano Zampini } 3007684f6988SStefano Zampini } 3008*5cbda25cSStefano Zampini /* free Neumann problem's matrix */ 3009*5cbda25cSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 3010304d26faSStefano Zampini PetscFunctionReturn(0); 3011304d26faSStefano Zampini } 3012304d26faSStefano Zampini 3013304d26faSStefano Zampini #undef __FUNCT__ 3014ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 301580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 3016674ae819SStefano Zampini { 3017674ae819SStefano Zampini PetscErrorCode ierr; 3018674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3019be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3020674ae819SStefano Zampini 3021674ae819SStefano Zampini PetscFunctionBegin; 3022df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 302380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 302420c7b377SStefano Zampini } 302580677318SStefano Zampini if (!pcbddc->switch_static) { 302680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 302780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 302880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 302920c7b377SStefano Zampini } 3030df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 303180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 303280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 303320c7b377SStefano Zampini } else { 3034df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3035be83ff47SStefano Zampini 3036df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3037df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 303820c7b377SStefano Zampini } 3039be83ff47SStefano Zampini } else { 304080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 304580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 304680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 304780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 304880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3049674ae819SStefano Zampini } 3050674ae819SStefano Zampini } 305172b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 305280677318SStefano Zampini if (applytranspose) { 305380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 305480677318SStefano Zampini } else { 305580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 305680677318SStefano Zampini } 3057be83ff47SStefano Zampini } else { 3058df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3059be83ff47SStefano Zampini 3060be83ff47SStefano Zampini if (applytranspose) { 3061df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3062be83ff47SStefano Zampini } else { 3063df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3064be83ff47SStefano Zampini } 3065be83ff47SStefano Zampini } 306680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 306780677318SStefano Zampini if (!pcbddc->switch_static) { 3068df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 306980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 307080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3071be83ff47SStefano Zampini } else { 3072df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3073be83ff47SStefano Zampini 3074df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3075df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3076be83ff47SStefano Zampini } 307780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 307880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 307980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 308080677318SStefano Zampini } 308180677318SStefano Zampini } else { 308280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 308680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 308780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 308880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 308980677318SStefano Zampini } 309080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 309380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3094674ae819SStefano Zampini } 3095674ae819SStefano Zampini PetscFunctionReturn(0); 3096674ae819SStefano Zampini } 3097674ae819SStefano Zampini 3098dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3099674ae819SStefano Zampini #undef __FUNCT__ 3100674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3101dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3102674ae819SStefano Zampini { 3103674ae819SStefano Zampini PetscErrorCode ierr; 3104674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3105674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3106674ae819SStefano Zampini const PetscScalar zero = 0.0; 3107674ae819SStefano Zampini 3108674ae819SStefano Zampini PetscFunctionBegin; 3109dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 31104fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3111dc359a40SStefano Zampini if (applytranspose) { 3112674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 31138eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3114dc359a40SStefano Zampini } else { 3115674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3116674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 311715aaf578SStefano Zampini } 31184fee134fSStefano Zampini } else { 31194fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 31204fee134fSStefano Zampini } 3121efc2fbd9SStefano Zampini 3122efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 31234f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3124efc2fbd9SStefano Zampini PetscScalar *array; 31254f1b2e48SStefano Zampini PetscInt j; 3126efc2fbd9SStefano Zampini 3127efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31284f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3129efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3130efc2fbd9SStefano Zampini } 3131efc2fbd9SStefano Zampini 313212edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 313312edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 313412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 313512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 313612edc857SStefano Zampini 31379f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 313812edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 313912edc857SStefano Zampini if (pcbddc->coarse_ksp) { 314051694757SStefano Zampini Mat coarse_mat; 3141964fefecSStefano Zampini Vec rhs,sol; 314251694757SStefano Zampini MatNullSpace nullsp; 3143964fefecSStefano Zampini 3144964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3145964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 314651694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 314751694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 314851694757SStefano Zampini if (nullsp) { 314951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 315051694757SStefano Zampini } 315112edc857SStefano Zampini if (applytranspose) { 3152964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 315312edc857SStefano Zampini } else { 3154964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 315512edc857SStefano Zampini } 315651694757SStefano Zampini if (nullsp) { 315751694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 315851694757SStefano Zampini } 315912edc857SStefano Zampini } 3160674ae819SStefano Zampini 3161674ae819SStefano Zampini /* Local solution on R nodes */ 31624fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 316380677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31649f00e9b4SStefano Zampini } 31659f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31669f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 316712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3168674ae819SStefano Zampini 31694fee134fSStefano Zampini /* Sum contributions from the two levels */ 31704fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3171dc359a40SStefano Zampini if (applytranspose) { 3172dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3173dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3174dc359a40SStefano Zampini } else { 3175674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31768eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3177dc359a40SStefano Zampini } 3178efc2fbd9SStefano Zampini /* store p0 */ 31794f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3180efc2fbd9SStefano Zampini PetscScalar *array; 31814f1b2e48SStefano Zampini PetscInt j; 3182efc2fbd9SStefano Zampini 3183efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31844f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3185efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3186efc2fbd9SStefano Zampini } 31874fee134fSStefano Zampini } else { /* expand the coarse solution */ 31884fee134fSStefano Zampini if (applytranspose) { 31894fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31904fee134fSStefano Zampini } else { 31914fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31924fee134fSStefano Zampini } 31934fee134fSStefano Zampini } 3194674ae819SStefano Zampini PetscFunctionReturn(0); 3195674ae819SStefano Zampini } 3196674ae819SStefano Zampini 3197674ae819SStefano Zampini #undef __FUNCT__ 3198674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 319912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3200674ae819SStefano Zampini { 3201674ae819SStefano Zampini PetscErrorCode ierr; 3202674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 320358da7f69SStefano Zampini PetscScalar *array; 320412edc857SStefano Zampini Vec from,to; 3205674ae819SStefano Zampini 3206674ae819SStefano Zampini PetscFunctionBegin; 320712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 320812edc857SStefano Zampini from = pcbddc->coarse_vec; 320912edc857SStefano Zampini to = pcbddc->vec1_P; 321012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 321112edc857SStefano Zampini Vec tvec; 321258da7f69SStefano Zampini 321358da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 321458da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 321512edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 321658da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 321758da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 321858da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 321912edc857SStefano Zampini } 322012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 322112edc857SStefano Zampini from = pcbddc->vec1_P; 322212edc857SStefano Zampini to = pcbddc->coarse_vec; 322312edc857SStefano Zampini } 322412edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3225674ae819SStefano Zampini PetscFunctionReturn(0); 3226674ae819SStefano Zampini } 3227674ae819SStefano Zampini 3228674ae819SStefano Zampini #undef __FUNCT__ 3229674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 323012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3231674ae819SStefano Zampini { 3232674ae819SStefano Zampini PetscErrorCode ierr; 3233674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 323458da7f69SStefano Zampini PetscScalar *array; 323512edc857SStefano Zampini Vec from,to; 3236674ae819SStefano Zampini 3237674ae819SStefano Zampini PetscFunctionBegin; 323812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 323912edc857SStefano Zampini from = pcbddc->coarse_vec; 324012edc857SStefano Zampini to = pcbddc->vec1_P; 324112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 324212edc857SStefano Zampini from = pcbddc->vec1_P; 324312edc857SStefano Zampini to = pcbddc->coarse_vec; 324412edc857SStefano Zampini } 324512edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 324612edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 324712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 324812edc857SStefano Zampini Vec tvec; 324958da7f69SStefano Zampini 325012edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 325158da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 325258da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 325358da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 325458da7f69SStefano Zampini } 325558da7f69SStefano Zampini } else { 325658da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 325758da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 325812edc857SStefano Zampini } 325912edc857SStefano Zampini } 3260674ae819SStefano Zampini PetscFunctionReturn(0); 3261674ae819SStefano Zampini } 3262674ae819SStefano Zampini 3263984c4197SStefano Zampini /* uncomment for testing purposes */ 3264984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3265674ae819SStefano Zampini #undef __FUNCT__ 3266674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3267674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3268674ae819SStefano Zampini { 3269674ae819SStefano Zampini PetscErrorCode ierr; 3270674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3271674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3272674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3273984c4197SStefano Zampini /* one and zero */ 3274984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3275984c4197SStefano Zampini /* space to store constraints and their local indices */ 32769162d606SStefano Zampini PetscScalar *constraints_data; 32779162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32789162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32799162d606SStefano Zampini PetscInt *constraints_n; 3280984c4197SStefano Zampini /* iterators */ 3281b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3282984c4197SStefano Zampini /* BLAS integers */ 3283e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3284e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3285c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3286727cdba6SStefano Zampini /* reuse */ 32870e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 32880e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3289984c4197SStefano Zampini /* change of basis */ 3290b3d85658SStefano Zampini PetscBool qr_needed; 32919162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3292984c4197SStefano Zampini /* auxiliary stuff */ 329364efe560SStefano Zampini PetscInt *nnz,*is_indices; 32948a0068c3SStefano Zampini PetscInt ncc; 3295984c4197SStefano Zampini /* some quantities */ 329645a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3297a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3298984c4197SStefano Zampini 3299674ae819SStefano Zampini PetscFunctionBegin; 33008e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 33018e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 33028e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3303088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3304088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 33050e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 33060e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 33070e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33080e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 33090e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3310088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3311cf5a6209SStefano Zampini 3312cf5a6209SStefano Zampini /* print some info */ 3313cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3314cf5a6209SStefano Zampini IS vertices; 3315cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3316cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3317cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3318cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 33191575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3320cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3321cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3322fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3323fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3324cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 33251575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3326cf5a6209SStefano Zampini } 3327cf5a6209SStefano Zampini 3328cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 33299162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3330cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3331cf5a6209SStefano Zampini const Vec *nearnullvecs; 3332cf5a6209SStefano Zampini Vec *localnearnullsp; 3333cf5a6209SStefano Zampini PetscScalar *array; 3334cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3335cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3336674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3337b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3338674ae819SStefano Zampini PetscScalar *work; 3339674ae819SStefano Zampini PetscReal *singular_vals; 3340674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3341674ae819SStefano Zampini PetscReal *rwork; 3342674ae819SStefano Zampini #endif 3343674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3344674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3345674ae819SStefano Zampini #else 3346964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3347964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3348674ae819SStefano Zampini #endif 3349674ae819SStefano Zampini 3350674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3351d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3352d06fc5fdSStefano Zampini /* free unneeded index sets */ 3353d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3354d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3355674ae819SStefano Zampini } 3356d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3357d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3358d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3359d06fc5fdSStefano Zampini } 3360d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3361d06fc5fdSStefano Zampini n_ISForEdges = 0; 3362d06fc5fdSStefano Zampini } 3363d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3364d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3365d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3366d06fc5fdSStefano Zampini } 3367d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3368d06fc5fdSStefano Zampini n_ISForFaces = 0; 3369d06fc5fdSStefano Zampini } 337070022509SStefano Zampini 337170022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 337270022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 337370022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 337470022509SStefano Zampini if (pcbddc->NullSpace) { 337570022509SStefano Zampini PetscBool tbool[2],gbool[2]; 337670022509SStefano Zampini 337770022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3378b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3379d06fc5fdSStefano Zampini if (!ISForEdges) { 3380d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3381d06fc5fdSStefano Zampini } 3382b8ffe317SStefano Zampini } 3383d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3384d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3385d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3386d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3387d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 338898a51de6SStefano Zampini } 338970022509SStefano Zampini #endif 339008122e43SStefano Zampini 3391674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3392674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3393674ae819SStefano Zampini if (nearnullsp) { 3394674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3395f4ddd8eeSStefano Zampini /* remove any stored info */ 3396f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3397f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3398f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3399f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3400f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3401473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3402f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3403f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3404f4ddd8eeSStefano Zampini } 3405984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3406984c4197SStefano Zampini nnsp_size = 0; 3407674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3408674ae819SStefano Zampini } 3409984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3410984c4197SStefano Zampini max_constraints = nnsp_size; 3411984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3412984c4197SStefano Zampini 3413674ae819SStefano Zampini /* 3414674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 34159162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 34169162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 34179162d606SStefano Zampini There can be multiple constraints per connected component 3418674ae819SStefano Zampini */ 3419674ae819SStefano Zampini n_vertices = 0; 3420674ae819SStefano Zampini if (ISForVertices) { 3421674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3422674ae819SStefano Zampini } 34239162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 34249162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 34259162d606SStefano Zampini 34269162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 34279162d606SStefano Zampini total_counts *= max_constraints; 3428674ae819SStefano Zampini total_counts += n_vertices; 34294641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34309162d606SStefano Zampini 3431674ae819SStefano Zampini total_counts = 0; 3432674ae819SStefano Zampini max_size_of_constraint = 0; 3433674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34349162d606SStefano Zampini IS used_is; 3435674ae819SStefano Zampini if (i<n_ISForEdges) { 34369162d606SStefano Zampini used_is = ISForEdges[i]; 3437674ae819SStefano Zampini } else { 34389162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3439674ae819SStefano Zampini } 34409162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3441674ae819SStefano Zampini total_counts += j; 3442674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3443674ae819SStefano Zampini } 34449162d606SStefano 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); 34459162d606SStefano Zampini 3446984c4197SStefano Zampini /* get local part of global near null space vectors */ 3447785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3448984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3449984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3450e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3451e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3452984c4197SStefano Zampini } 3453674ae819SStefano Zampini 3454242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3455242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3456a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3457242a89d7SStefano Zampini 3458984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3459a773dcb8SStefano Zampini if (!skip_lapack) { 3460674ae819SStefano Zampini PetscScalar temp_work; 3461911cabfeSStefano Zampini 3462674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3463984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3464785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3465785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3466785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3467674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3468785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3469674ae819SStefano Zampini #endif 3470674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3471c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3472c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3473674ae819SStefano Zampini lwork = -1; 3474674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3475674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3476c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3477674ae819SStefano Zampini #else 3478c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3479674ae819SStefano Zampini #endif 3480674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3481984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3482674ae819SStefano Zampini #else /* on missing GESVD */ 3483674ae819SStefano Zampini /* SVD */ 3484674ae819SStefano Zampini PetscInt max_n,min_n; 3485674ae819SStefano Zampini max_n = max_size_of_constraint; 3486984c4197SStefano Zampini min_n = max_constraints; 3487984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3488674ae819SStefano Zampini min_n = max_size_of_constraint; 3489984c4197SStefano Zampini max_n = max_constraints; 3490674ae819SStefano Zampini } 3491785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3492674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3493785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3494674ae819SStefano Zampini #endif 3495674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3496674ae819SStefano Zampini lwork = -1; 3497e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3498e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3499b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3500674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3501674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35029162d606SStefano 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)); 3503674ae819SStefano Zampini #else 35049162d606SStefano 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)); 3505674ae819SStefano Zampini #endif 3506674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3507984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3508984c4197SStefano Zampini #endif /* on missing GESVD */ 3509674ae819SStefano Zampini /* Allocate optimal workspace */ 3510674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3511854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3512674ae819SStefano Zampini } 3513674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3514674ae819SStefano Zampini total_counts = 0; 35159162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 35169162d606SStefano Zampini constraints_data_ptr[0] = 0; 3517674ae819SStefano Zampini /* vertices */ 35189162d606SStefano Zampini if (n_vertices) { 3519674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 35209162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3521674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 35229162d606SStefano Zampini constraints_n[total_counts] = 1; 35239162d606SStefano Zampini constraints_data[total_counts] = 1.0; 35249162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 35259162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3526674ae819SStefano Zampini total_counts++; 3527674ae819SStefano Zampini } 3528674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3529674ae819SStefano Zampini n_vertices = total_counts; 3530674ae819SStefano Zampini } 3531984c4197SStefano Zampini 3532674ae819SStefano Zampini /* edges and faces */ 35339162d606SStefano Zampini total_counts_cc = total_counts; 3534911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35359162d606SStefano Zampini IS used_is; 35369162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35379162d606SStefano Zampini 3538911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35399162d606SStefano Zampini used_is = ISForEdges[ncc]; 3540984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3541674ae819SStefano Zampini } else { 35429162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3543984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3544674ae819SStefano Zampini } 3545674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35469162d606SStefano Zampini 35479162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35489162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3549984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3550984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3551674ae819SStefano Zampini if (nnsp_has_cnst) { 35525b08dc53SStefano Zampini PetscScalar quad_value; 35539162d606SStefano Zampini 35549162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35559162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35569162d606SStefano Zampini 3557a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3558674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3559a773dcb8SStefano Zampini } else { 3560a773dcb8SStefano Zampini quad_value = 1.0; 3561a773dcb8SStefano Zampini } 3562674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35639162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3564674ae819SStefano Zampini } 35659162d606SStefano Zampini temp_constraints++; 3566674ae819SStefano Zampini total_counts++; 3567674ae819SStefano Zampini } 3568674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3569984c4197SStefano Zampini PetscReal real_value; 35709162d606SStefano Zampini PetscScalar *ptr_to_data; 35719162d606SStefano Zampini 3572984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35739162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3574674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35759162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3576674ae819SStefano Zampini } 3577984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3578984c4197SStefano Zampini /* check if array is null on the connected component */ 3579e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35809162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35815b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3582674ae819SStefano Zampini temp_constraints++; 3583674ae819SStefano Zampini total_counts++; 35849162d606SStefano Zampini if (!idxs_copied) { 35859162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35869162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3587674ae819SStefano Zampini } 3588674ae819SStefano Zampini } 35899162d606SStefano Zampini } 35909162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 359145a1bb75SStefano Zampini valid_constraints = temp_constraints; 3592eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3593a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 35949162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 35959162d606SStefano Zampini 35969162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3597a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35989162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3599a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 36009162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3601a773dcb8SStefano Zampini } else { /* perform SVD */ 3602984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 36039162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3604674ae819SStefano Zampini 3605674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3606984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3607984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3608984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3609984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3610984c4197SStefano Zampini from that computed using LAPACKgesvd 3611984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3612984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3613984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3614674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3615e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3616984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3617674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3618674ae819SStefano Zampini for (k=0;k<j+1;k++) { 36199162d606SStefano 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)); 3620674ae819SStefano Zampini } 3621674ae819SStefano Zampini } 3622e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3623e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3624e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3625674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3626c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3627674ae819SStefano Zampini #else 3628c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3629674ae819SStefano Zampini #endif 3630674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3631984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3632984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3633674ae819SStefano Zampini j = 0; 3634984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3635674ae819SStefano Zampini total_counts = total_counts-j; 363645a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3637e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3638c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3639c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3640c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3641c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3642c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3643c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3644674ae819SStefano Zampini if (j<temp_constraints) { 3645984c4197SStefano Zampini PetscInt ii; 3646984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3647674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36489162d606SStefano 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)); 3649674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3650984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3651674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36529162d606SStefano 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]; 3653674ae819SStefano Zampini } 3654674ae819SStefano Zampini } 3655674ae819SStefano Zampini } 3656674ae819SStefano Zampini #else /* on missing GESVD */ 3657e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3658e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3659b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3660674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3661674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36629162d606SStefano 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)); 3663674ae819SStefano Zampini #else 36649162d606SStefano 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)); 3665674ae819SStefano Zampini #endif 3666984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3667674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3668984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3669e310c8b4SStefano Zampini k = temp_constraints; 3670e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3671674ae819SStefano Zampini j = 0; 3672e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 367345a1bb75SStefano Zampini valid_constraints = k-j; 3674911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3675984c4197SStefano Zampini #endif /* on missing GESVD */ 3676674ae819SStefano Zampini } 3677a773dcb8SStefano Zampini } 36789162d606SStefano Zampini /* update pointers information */ 36799162d606SStefano Zampini if (valid_constraints) { 36809162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36819162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36829162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36839162d606SStefano Zampini /* set change_of_basis flag */ 368445a1bb75SStefano Zampini if (boolforchange) { 3685b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 36869162d606SStefano Zampini } 3687b3d85658SStefano Zampini total_counts_cc++; 368845a1bb75SStefano Zampini } 368945a1bb75SStefano Zampini } 3690984c4197SStefano Zampini /* free workspace */ 36918f1c130eSStefano Zampini if (!skip_lapack) { 3692984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3693984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3694984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3695984c4197SStefano Zampini #endif 3696984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3697984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3698984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3699984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3700984c4197SStefano Zampini #endif 3701984c4197SStefano Zampini } 3702984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3703984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3704984c4197SStefano Zampini } 3705984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3706cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3707cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3708cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3709cf5a6209SStefano Zampini } 3710cf5a6209SStefano Zampini if (n_ISForFaces) { 3711cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3712cf5a6209SStefano Zampini } 3713cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3714cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3715cf5a6209SStefano Zampini } 3716cf5a6209SStefano Zampini if (n_ISForEdges) { 3717cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3718cf5a6209SStefano Zampini } 3719cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 372008122e43SStefano Zampini } else { 372108122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3722984c4197SStefano Zampini 372308122e43SStefano Zampini total_counts = 0; 372408122e43SStefano Zampini n_vertices = 0; 3725d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3726d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 372708122e43SStefano Zampini } 372808122e43SStefano Zampini max_constraints = 0; 37299162d606SStefano Zampini total_counts_cc = 0; 373008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 373108122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37329162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 373308122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 373408122e43SStefano Zampini } 37359162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37369162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37379162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37389162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 373974d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37409162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37419162d606SStefano Zampini total_counts_cc = 0; 37429162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37439162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37449162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 374508122e43SStefano Zampini } 374608122e43SStefano Zampini } 37479162d606SStefano Zampini #if 0 37489162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37499162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37509162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37519162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37529162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37539162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37549162d606SStefano Zampini } 37559162d606SStefano Zampini printf("\n"); 37569162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37579162d606SStefano Zampini } 37581b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37598bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37601b968477SStefano Zampini } 37611b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37628bec7fa6SStefano 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]); 37631b968477SStefano Zampini } 376408122e43SStefano Zampini #endif 376508122e43SStefano Zampini 37668bec7fa6SStefano Zampini max_size_of_constraint = 0; 37679162d606SStefano 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]); 37689162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 376908122e43SStefano Zampini /* Change of basis */ 3770b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 377108122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 377208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 377308122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3774b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 377508122e43SStefano Zampini } 377608122e43SStefano Zampini } 377708122e43SStefano Zampini } 377808122e43SStefano Zampini } 3779984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37804f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 378108122e43SStefano Zampini 37829162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37839162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 37849162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 37859162d606SStefano 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); 378608122e43SStefano Zampini } 3787674ae819SStefano Zampini 3788674ae819SStefano Zampini /* Create constraint matrix */ 3789674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 379016f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3791984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3792984c4197SStefano Zampini 3793984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3794a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3795a717540cSStefano Zampini qr_needed = PETSC_FALSE; 379674d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3797984c4197SStefano Zampini total_primal_vertices=0; 3798b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 37999162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 38009162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 380172b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 38029162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3803b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 380464efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 38059162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38069162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3807a717540cSStefano Zampini } 3808b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 380974d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3810a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3811a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3812a717540cSStefano Zampini } 3813fa434743SStefano Zampini } else { 3814b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3815fa434743SStefano Zampini } 3816a717540cSStefano Zampini } 3817b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3818b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3819674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 382070022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3821b3d85658SStefano Zampini 38224f1b2e48SStefano 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); 38230e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 38240e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3825984c4197SStefano Zampini 3826984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 382774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3828785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3829984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 383074d5cdf7SStefano Zampini 3831984c4197SStefano Zampini j = total_primal_vertices; 383274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3833b3d85658SStefano Zampini cum = total_primal_vertices; 38349162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38354641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3836b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3837b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3838b3d85658SStefano Zampini cum++; 38399162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 384074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 384174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 384274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 384374d5cdf7SStefano Zampini } 38449162d606SStefano Zampini j += constraints_n[i]; 3845674ae819SStefano Zampini } 3846674ae819SStefano Zampini } 3847674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3848674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3849088faed8SStefano Zampini 3850674ae819SStefano Zampini /* set values in constraint matrix */ 3851984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38520e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3853674ae819SStefano Zampini } 3854984c4197SStefano Zampini total_counts = total_primal_vertices; 38559162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38564641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38579162d606SStefano Zampini PetscInt *cols; 38589162d606SStefano Zampini 38599162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38609162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38619162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38629162d606SStefano Zampini PetscInt row = total_counts+k; 38639162d606SStefano Zampini PetscScalar *vals; 38649162d606SStefano Zampini 38659162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38669162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38679162d606SStefano Zampini } 38689162d606SStefano Zampini total_counts += constraints_n[i]; 3869674ae819SStefano Zampini } 3870674ae819SStefano Zampini } 3871674ae819SStefano Zampini /* assembling */ 3872674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3873674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3874088faed8SStefano Zampini 3875984c4197SStefano Zampini /* 387645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3877984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3878984c4197SStefano Zampini */ 3879674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3880674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3881026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3882984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3883984c4197SStefano Zampini /* working stuff for GEQRF */ 388481d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3885984c4197SStefano Zampini PetscBLASInt lqr_work; 3886984c4197SStefano Zampini /* working stuff for UNGQR */ 3887984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3888984c4197SStefano Zampini PetscBLASInt lgqr_work; 3889984c4197SStefano Zampini /* working stuff for TRTRS */ 3890984c4197SStefano Zampini PetscScalar *trs_rhs; 38913f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3892984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3893984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3894984c4197SStefano Zampini PetscScalar *start_vals; 3895984c4197SStefano Zampini /* working stuff for values insertion */ 38964641a718SStefano Zampini PetscBT is_primal; 389764efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3898906d46d4SStefano Zampini /* matrix sizes */ 3899906d46d4SStefano Zampini PetscInt global_size,local_size; 3900906d46d4SStefano Zampini /* temporary change of basis */ 3901906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3902cf5a6209SStefano Zampini /* extra space for debugging */ 3903cf5a6209SStefano Zampini PetscScalar *dbg_work; 3904984c4197SStefano Zampini 3905906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3906906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 390716f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3908bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3909906d46d4SStefano Zampini /* nonzeros for local mat */ 3910bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3911bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 39129162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3913a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 39149162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3915a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 39169162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3917a717540cSStefano Zampini } else { 39189162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 39199162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3920a717540cSStefano Zampini } 3921a717540cSStefano Zampini } 3922a717540cSStefano Zampini } 3923906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3924bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3925a717540cSStefano Zampini /* Set initial identity in the matrix */ 3926bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3927906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3928a717540cSStefano Zampini } 3929a717540cSStefano Zampini 3930a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3931a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3932a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3933a717540cSStefano Zampini } 3934a717540cSStefano Zampini 3935a717540cSStefano Zampini 3936a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3937a717540cSStefano Zampini /* 3938a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3939a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3940a717540cSStefano Zampini 3941a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3942a717540cSStefano Zampini 3943a6b551f4SStefano 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) 3944a6b551f4SStefano Zampini 3945a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3946a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3947a717540cSStefano Zampini | ... | 3948a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3949a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3950a717540cSStefano Zampini 3951a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3952a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3953a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3954a6b551f4SStefano Zampini 3955a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3956a717540cSStefano Zampini */ 3957a717540cSStefano Zampini if (qr_needed) { 3958984c4197SStefano Zampini /* space to store Q */ 3959854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3960984c4197SStefano Zampini /* first we issue queries for optimal work */ 39613f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39623f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 39633f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3964984c4197SStefano Zampini lqr_work = -1; 39653f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3966984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3967984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3968785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3969984c4197SStefano Zampini lgqr_work = -1; 39703f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39713f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 39723f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 39733f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39743f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 39753f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3976984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3977984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3978785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3979984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3980785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3981984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3982785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3983a717540cSStefano Zampini /* allocating workspace for check */ 3984a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3985cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3986a717540cSStefano Zampini } 3987a717540cSStefano Zampini } 3988984c4197SStefano Zampini /* array to store whether a node is primal or not */ 39894641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3990473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 39910e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 399239e2fb2aSStefano Zampini if (i != total_primal_vertices) { 399339e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 39944641a718SStefano Zampini } 399539e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 399639e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 399739e2fb2aSStefano Zampini } 399839e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3999984c4197SStefano Zampini 4000a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 40019162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 40029162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 40034641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 4004984c4197SStefano Zampini /* get constraint info */ 40059162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 4006984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 4007984c4197SStefano Zampini 4008984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40099162d606SStefano 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); 4010674ae819SStefano Zampini } 4011984c4197SStefano Zampini 4012fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 4013a717540cSStefano Zampini 4014a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 4015a717540cSStefano Zampini if (pcbddc->dbg_flag) { 40169162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4017a717540cSStefano Zampini } 4018984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 40199162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4020984c4197SStefano Zampini 4021984c4197SStefano Zampini /* compute QR decomposition of constraints */ 40223f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40233f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40243f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4025674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40263f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 4027984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 4028674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4029984c4197SStefano Zampini 4030984c4197SStefano Zampini /* explictly compute R^-T */ 4031984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4032984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40333f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40343f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40353f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4037984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40383f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4039984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4040984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4041984c4197SStefano Zampini 4042a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40433f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40443f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40453f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40463f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4047984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40483f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4049984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4050984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4051984c4197SStefano Zampini 4052984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4053984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4054984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40553f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40563f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40573f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40583f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40593f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40603f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4061984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40629162d606SStefano 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)); 4063984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40649162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4065984c4197SStefano Zampini 4066984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 40679162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4068984c4197SStefano Zampini /* insert cols for primal dofs */ 4069984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4070984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 40719162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4072906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4073984c4197SStefano Zampini } 4074984c4197SStefano Zampini /* insert cols for dual dofs */ 4075984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 40769162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4077984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 40789162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4079906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4080984c4197SStefano Zampini j++; 4081674ae819SStefano Zampini } 4082674ae819SStefano Zampini } 4083984c4197SStefano Zampini 4084984c4197SStefano Zampini /* check change of basis */ 4085984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4086984c4197SStefano Zampini PetscInt ii,jj; 4087984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4088c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4089c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4090c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4091c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4092c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4093c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4094984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4095cf5a6209SStefano 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)); 4096984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4097984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4098984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4099cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4100cf5a6209SStefano 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; 4101674ae819SStefano Zampini } 4102674ae819SStefano Zampini } 4103984c4197SStefano Zampini if (!valid_qr) { 410422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4105984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4106984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4107cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4108cf5a6209SStefano 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])); 4109674ae819SStefano Zampini } 4110cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4111cf5a6209SStefano 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])); 4112984c4197SStefano Zampini } 4113984c4197SStefano Zampini } 4114984c4197SStefano Zampini } 4115674ae819SStefano Zampini } else { 411622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4117674ae819SStefano Zampini } 4118674ae819SStefano Zampini } 4119a717540cSStefano Zampini } else { /* simple transformation block */ 4120a717540cSStefano Zampini PetscInt row,col; 4121a6b551f4SStefano Zampini PetscScalar val,norm; 4122a6b551f4SStefano Zampini 4123a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 41249162d606SStefano 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)); 4125a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 41269162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 41279162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4128bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 41299162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4130906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41319162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4132a717540cSStefano Zampini } else { 4133a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41349162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4135a717540cSStefano Zampini if (row != col) { 41369162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4137a717540cSStefano Zampini } else { 41389162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4139a717540cSStefano Zampini } 4140906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4141a717540cSStefano Zampini } 4142a717540cSStefano Zampini } 4143a717540cSStefano Zampini } 414498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 414522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4146a717540cSStefano Zampini } 4147674ae819SStefano Zampini } 4148984c4197SStefano Zampini } else { 4149984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41509162d606SStefano 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); 4151674ae819SStefano Zampini } 4152674ae819SStefano Zampini } 4153674ae819SStefano Zampini } 4154a717540cSStefano Zampini 4155a717540cSStefano Zampini /* free workspace */ 4156a717540cSStefano Zampini if (qr_needed) { 4157984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4158cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4159984c4197SStefano Zampini } 4160984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4161984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4162984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4163984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4164984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4165674ae819SStefano Zampini } 4166a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4167906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4168906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4169906d46d4SStefano Zampini 4170906d46d4SStefano Zampini /* assembling of global change of variable */ 417188c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4172bbb9e6c6SStefano Zampini Mat tmat; 417316f15bc4SStefano Zampini PetscInt bs; 417416f15bc4SStefano Zampini 4175906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4176906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4177bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4178bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4179bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4180bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 418116f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 418216f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4183906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4184bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4185bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4186bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4187bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4188bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4189e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4190e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4191bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4192bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 419388c03ad3SStefano Zampini 4194906d46d4SStefano Zampini /* check */ 4195906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4196906d46d4SStefano Zampini PetscReal error; 4197906d46d4SStefano Zampini Vec x,x_change; 4198906d46d4SStefano Zampini 4199906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4200906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4201906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4202906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4203e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4204e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4205bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4206e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4207e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4208906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4209906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4210906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4211906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4212bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4213906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4214906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4215906d46d4SStefano Zampini } 4216b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4217b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4218b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4219bf3a8328SStefano Zampini 4220bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4221bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4222bf3a8328SStefano Zampini } 4223b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4224ac632422SStefano Zampini Mat S_new,tmat; 4225bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4226bbb9e6c6SStefano Zampini 4227bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 42286816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4229bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4230bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4231bf3a8328SStefano Zampini IS is_V; 4232b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4233b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4234b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4235b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4236b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4237bf3a8328SStefano Zampini } 4238bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4239ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4240b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4241ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4242bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4243bf3a8328SStefano Zampini const PetscScalar *array; 4244bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4245bf3a8328SStefano Zampini PetscInt i,n_V; 4246bf3a8328SStefano Zampini 4247b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4248b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4249b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4250b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4251b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4252b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4253b087196eSStefano Zampini PetscScalar val; 4254b087196eSStefano Zampini PetscInt idx; 4255b087196eSStefano Zampini 4256b087196eSStefano Zampini idx = idxs_V[i]; 4257b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4258b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4259b087196eSStefano Zampini } 4260b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4261b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4262bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4263bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4264bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4265bf3a8328SStefano Zampini } 4266ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4267ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4268ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4269ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4270b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4271ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4272bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4273b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4274bf3a8328SStefano Zampini } 4275ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4276ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4277ac632422SStefano Zampini } 4278b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 427988c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4280b96c3477SStefano Zampini } 4281c9db6a07SStefano Zampini /* destroy any change of basis context in sub_schurs */ 4282c9db6a07SStefano Zampini if (sub_schurs->change) { 4283c9db6a07SStefano Zampini PetscInt i; 4284c9db6a07SStefano Zampini 4285c9db6a07SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 4286c9db6a07SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 4287c9db6a07SStefano Zampini } 4288c9db6a07SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 4289c9db6a07SStefano Zampini } 4290b96c3477SStefano Zampini } 4291906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 429272b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 429372b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 429472b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 429572b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 429672b8c272SStefano Zampini } 4297906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4298b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4299b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4300b9b85e73SStefano Zampini } 4301906d46d4SStefano Zampini 4302906d46d4SStefano Zampini /* set up change of basis context */ 4303906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4304906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4305906d46d4SStefano Zampini 4306906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4307906d46d4SStefano Zampini PetscInt global_size,local_size; 4308906d46d4SStefano Zampini 4309906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4310906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4311906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4312906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4313906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4314906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4315906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4316906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4317906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4318906d46d4SStefano Zampini } else { 4319906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4320906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4321906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4322906d46d4SStefano Zampini } 4323906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4324906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4325906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4326906d46d4SStefano Zampini } else { 4327906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4328906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4329906d46d4SStefano Zampini } 4330906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4331906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4332906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4333906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4334a8661815SStefano Zampini } else { 4335a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4336b9b85e73SStefano Zampini } 4337a717540cSStefano Zampini 433872b8c272SStefano Zampini if (!pcbddc->fake_change) { 43394f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 43404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 43414f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 43424f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4343019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4344019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4345019a44ceSStefano Zampini pcbddc->local_primal_size++; 4346019a44ceSStefano Zampini } 4347019a44ceSStefano Zampini 4348019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4349727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4350727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 43519f47a83aSStefano 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); 4352c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 43530e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 43549f47a83aSStefano 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); 4355727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4356727cdba6SStefano Zampini } 43570e6343abSStefano Zampini } 4358727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4359727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 436072b8c272SStefano Zampini } 436172b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4362727cdba6SStefano Zampini 4363a717540cSStefano Zampini /* flush dbg viewer */ 4364b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4365b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4366b8ffe317SStefano Zampini } 4367a717540cSStefano Zampini 4368e310c8b4SStefano Zampini /* free workspace */ 4369a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 43704641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 437108122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 43729162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 43739162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 437408122e43SStefano Zampini } else { 43759162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 43769162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 43779162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 437808122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 437908122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 43809162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 43819162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 438208122e43SStefano Zampini } 4383674ae819SStefano Zampini PetscFunctionReturn(0); 4384674ae819SStefano Zampini } 4385674ae819SStefano Zampini 4386674ae819SStefano Zampini #undef __FUNCT__ 4387674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4388674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4389674ae819SStefano Zampini { 4390674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4391674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4392674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 439314f95afaSStefano Zampini PetscInt ierr,i,N; 4394674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4395674ae819SStefano Zampini 4396674ae819SStefano Zampini PetscFunctionBegin; 43978e61c736SStefano Zampini /* Reset previously computed graph */ 43988e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4399674ae819SStefano Zampini /* Init local Graph struct */ 44007fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 44013bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4402674ae819SStefano Zampini 4403575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 44045099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 44055099eff2SStefano 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); 4406575ad6abSStefano Zampini } 44079577ea80SStefano Zampini 4408674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4409d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 44104d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 44114d379d7bSStefano Zampini PetscInt nvtxs; 4412e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4413674ae819SStefano Zampini 44142fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 44152fffb893SStefano Zampini if (flg_row) { 44164d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4417b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 44182fffb893SStefano Zampini } 44192fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4420674ae819SStefano Zampini } 44219b28b941SStefano Zampini if (pcbddc->dbg_flag) { 44229b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4423674ae819SStefano Zampini } 4424674ae819SStefano Zampini 442563602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 442663602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 442763602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 442895ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 442963602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4430e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 443163602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4432674ae819SStefano Zampini } 443363602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 443463602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 443563602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4436674ae819SStefano Zampini } 443763602bcaSStefano Zampini } else { 443863602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 443963602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4440854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 444163602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 444263602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 444363602bcaSStefano Zampini } 444463602bcaSStefano Zampini } 4445674ae819SStefano Zampini } 4446674ae819SStefano Zampini 4447674ae819SStefano Zampini /* Setup of Graph */ 4448785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4449e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4450785d1243SStefano Zampini } 4451785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4452e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4453785d1243SStefano Zampini } 445430368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 445530368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 445630368db7SStefano Zampini } 445714f95afaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,pcbddc->vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4458674ae819SStefano Zampini 44594f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 44604f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 44614f1b2e48SStefano Zampini PetscInt *local_subs; 44624f1b2e48SStefano Zampini 44634f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 44644f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 44654f1b2e48SStefano Zampini const PetscInt *idxs; 44664f1b2e48SStefano Zampini PetscInt nl,j; 44674f1b2e48SStefano Zampini 44684f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 44694f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44704f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 44714f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 44724f1b2e48SStefano Zampini } 44734f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44744f1b2e48SStefano Zampini } 44754f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 44764f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 44774f1b2e48SStefano Zampini } 44784f1b2e48SStefano Zampini 4479674ae819SStefano Zampini /* Graph's connected components analysis */ 4480674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4481674ae819SStefano Zampini 4482674ae819SStefano Zampini /* print some info to stdout */ 4483674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4484302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4485674ae819SStefano Zampini } 4486674ae819SStefano Zampini PetscFunctionReturn(0); 4487674ae819SStefano Zampini } 4488674ae819SStefano Zampini 4489dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4490674ae819SStefano Zampini #undef __FUNCT__ 4491674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4492dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4493674ae819SStefano Zampini { 4494dc456d91SStefano Zampini PetscSF sf; 4495dc456d91SStefano Zampini PetscLayout map; 4496dc456d91SStefano Zampini const PetscInt *idxs; 4497dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4498dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4499dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4500dc456d91SStefano Zampini PetscMPIInt commsize; 4501674ae819SStefano Zampini PetscBool first_found; 4502674ae819SStefano Zampini PetscErrorCode ierr; 4503674ae819SStefano Zampini 4504674ae819SStefano Zampini PetscFunctionBegin; 4505dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4506dc456d91SStefano Zampini if (subset_mult) { 4507dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4508dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4509dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4510674ae819SStefano Zampini } 4511dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4512dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4513dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4514dc456d91SStefano Zampini for (i=0;i<n;i++) { 4515dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4516dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4517674ae819SStefano Zampini } 4518dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4519dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4520dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4521dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4522dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4523dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4524dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4525dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4526dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4527dc456d91SStefano Zampini 4528dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4529dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4530dc456d91SStefano Zampini if (subset_mult) { 4531dc456d91SStefano Zampini const PetscInt* idxs_mult; 4532dc456d91SStefano Zampini 4533dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4534dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4535dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4536674ae819SStefano Zampini } else { 4537dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4538674ae819SStefano Zampini } 4539dc456d91SStefano Zampini /* local size of new subset */ 4540dc456d91SStefano Zampini n_n = 0; 4541dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4542dc456d91SStefano Zampini 4543dc456d91SStefano Zampini /* global indexes in layout */ 4544dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4545dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4546dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4547dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4548dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4549dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4550dc456d91SStefano Zampini 4551dc456d91SStefano Zampini /* reduce from leaves to roots */ 4552dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 455364a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 455464a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4555dc456d91SStefano Zampini 4556dc456d91SStefano Zampini /* count indexes in local part of layout */ 4557674ae819SStefano Zampini nlocals = 0; 4558674ae819SStefano Zampini first_index = -1; 4559674ae819SStefano Zampini first_found = PETSC_FALSE; 4560dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4561dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4562674ae819SStefano Zampini first_found = PETSC_TRUE; 4563674ae819SStefano Zampini first_index = i; 4564674ae819SStefano Zampini } 4565dc456d91SStefano Zampini nlocals += root_data[i]; 4566674ae819SStefano Zampini } 4567dc456d91SStefano Zampini 4568dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 45695fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4570dc456d91SStefano Zampini start = 0; 457164a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45725fa240b1SStefano Zampini #else 457364a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45745fa240b1SStefano Zampini start = start-nlocals; 45755fa240b1SStefano Zampini #endif 45765fa240b1SStefano Zampini 4577dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4578dc456d91SStefano Zampini *N_n = start + nlocals; 4579dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4580dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4581674ae819SStefano Zampini } 45825fa240b1SStefano Zampini 45835fa240b1SStefano Zampini /* adapt root data with cumulative */ 4584674ae819SStefano Zampini if (first_found) { 4585dc456d91SStefano Zampini PetscInt old_index; 4586dc456d91SStefano Zampini 4587dc456d91SStefano Zampini root_data[first_index] += start; 4588674ae819SStefano Zampini old_index = first_index; 4589dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4590dc456d91SStefano Zampini if (root_data[i]) { 4591dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4592674ae819SStefano Zampini old_index = i; 4593674ae819SStefano Zampini } 4594674ae819SStefano Zampini } 4595674ae819SStefano Zampini } 4596dc456d91SStefano Zampini 4597dc456d91SStefano Zampini /* from roots to leaves */ 4598dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4599dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4600dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4601dc456d91SStefano Zampini 4602dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4603dc456d91SStefano Zampini if (subset_mult) { 4604dc456d91SStefano Zampini const PetscInt* idxs_mult; 4605dc456d91SStefano Zampini PetscInt cum; 4606dc456d91SStefano Zampini 4607dc456d91SStefano Zampini cum = 0; 4608dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4609dc456d91SStefano Zampini for (i=0;i<n;i++) { 4610dc456d91SStefano Zampini PetscInt j; 4611dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4612674ae819SStefano Zampini } 4613dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4614674ae819SStefano Zampini } else { 4615dc456d91SStefano Zampini for (i=0;i<n;i++) { 4616dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4617674ae819SStefano Zampini } 4618674ae819SStefano Zampini } 4619dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4620dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4621674ae819SStefano Zampini PetscFunctionReturn(0); 4622674ae819SStefano Zampini } 46239a7d3425SStefano Zampini 46249a7d3425SStefano Zampini #undef __FUNCT__ 46259a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 46269a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 46279a7d3425SStefano Zampini { 46289a7d3425SStefano Zampini PetscInt i,j; 46299a7d3425SStefano Zampini PetscScalar *alphas; 46309a7d3425SStefano Zampini PetscErrorCode ierr; 46319a7d3425SStefano Zampini 46329a7d3425SStefano Zampini PetscFunctionBegin; 46339a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4634785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 46359a7d3425SStefano Zampini for (i=0;i<n;i++) { 46369a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 46379a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 46389a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 46399a7d3425SStefano Zampini } 46409a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 46419a7d3425SStefano Zampini PetscFunctionReturn(0); 46429a7d3425SStefano Zampini } 46439a7d3425SStefano Zampini 4644e7931f94SStefano Zampini #undef __FUNCT__ 464570cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4646b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4647e7931f94SStefano Zampini { 464852e5ac9dSStefano Zampini IS ranks_send_to; 4649e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4650e7931f94SStefano Zampini PetscMPIInt size,rank,color; 465152e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 465252e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 46533837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 46542b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4655e7931f94SStefano Zampini PetscSubcomm subcomm; 465652e5ac9dSStefano Zampini PetscErrorCode ierr; 4657a57a6d2fSStefano Zampini 4658e7931f94SStefano Zampini PetscFunctionBegin; 46592b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 46602b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 46612b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4662e7931f94SStefano Zampini 4663e7931f94SStefano Zampini /* Get info on mapping */ 46643bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 46653bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4666e7931f94SStefano Zampini 4667e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4668785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4669e7931f94SStefano Zampini xadj[0] = 0; 4670e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4671785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4672785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4673e7931f94SStefano Zampini 46742b510759SStefano Zampini if (threshold) { 4675d023bfaeSStefano Zampini PetscInt xadj_count = 0; 46762b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4677d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4678d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4679d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4680d023bfaeSStefano Zampini xadj_count++; 4681e7931f94SStefano Zampini } 4682e7931f94SStefano Zampini } 4683d023bfaeSStefano Zampini xadj[1] = xadj_count; 4684c8587f34SStefano Zampini } else { 4685e7931f94SStefano Zampini if (xadj[1]) { 4686e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4687e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4688c8587f34SStefano Zampini } 4689e7931f94SStefano Zampini } 46903bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4691e7931f94SStefano Zampini if (use_square) { 4692e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4693e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4694e7931f94SStefano Zampini } 4695e7931f94SStefano Zampini } 4696e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4697e7931f94SStefano Zampini 46983837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4699e7931f94SStefano Zampini 4700e7931f94SStefano Zampini /* 4701e7931f94SStefano Zampini Restrict work on active processes only. 4702e7931f94SStefano Zampini */ 4703e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4704e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4705e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 47062b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4707d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4708e7931f94SStefano Zampini if (color) { 4709e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4710e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4711e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4712c8587f34SStefano Zampini } else { 471352e5ac9dSStefano Zampini Mat subdomain_adj; 471452e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 471552e5ac9dSStefano Zampini MatPartitioning partitioner; 471652e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 471752e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4718b0c7d250SStefano Zampini PetscBool aggregate; 4719b0c7d250SStefano Zampini 4720306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4721785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4722e7931f94SStefano Zampini prank = rank; 4723306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 47248002ef2cSStefano Zampini /* 4725e7931f94SStefano Zampini for (i=0;i<size;i++) { 4726e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4727c8587f34SStefano Zampini } 47288002ef2cSStefano Zampini */ 4729e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4730e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4731c8587f34SStefano Zampini } 4732e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4733b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4734b0c7d250SStefano Zampini if (aggregate) { 4735b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4736b0c7d250SStefano Zampini PetscMPIInt nrank; 4737b0c7d250SStefano Zampini PetscScalar *vals; 4738b0c7d250SStefano Zampini 4739b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4740b0c7d250SStefano Zampini lrows = 0; 4741b0c7d250SStefano Zampini if (nrank<redprocs) { 4742b0c7d250SStefano Zampini lrows = size/redprocs; 4743b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4744b0c7d250SStefano Zampini } 47455fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4746b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4747b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4748b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4749b0c7d250SStefano Zampini row = nrank; 4750b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4751b0c7d250SStefano Zampini cols = adjncy; 4752b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4753b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4754b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4755b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4756b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 475752e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 475852e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 475952e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4760b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4761b0c7d250SStefano Zampini } else { 4762306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4763b0c7d250SStefano Zampini } 476422b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4765e7931f94SStefano Zampini 4766e7931f94SStefano Zampini /* Partition */ 4767306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4768e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4769e7931f94SStefano Zampini if (use_vwgt) { 47703837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4771e7931f94SStefano Zampini v_wgt[0] = local_size; 4772e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4773c8587f34SStefano Zampini } 477428143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 477528143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4776e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4777e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 477822b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4779e7931f94SStefano Zampini 478052e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 478152e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 478252e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 478352e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4784b0c7d250SStefano Zampini if (!redprocs) { 4785b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 478628143c3dSStefano Zampini } else { 4787b0c7d250SStefano Zampini PetscInt idxs[1]; 4788b0c7d250SStefano Zampini PetscMPIInt tag; 4789b0c7d250SStefano Zampini MPI_Request *reqs; 4790b0c7d250SStefano Zampini 4791b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4792b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4793b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4794b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 479528143c3dSStefano Zampini } 4796b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4797b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4798b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4799b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4800e7931f94SStefano Zampini } 480152e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4802e7931f94SStefano Zampini /* clean up */ 4803e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 480452e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4805e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4806e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4807e7931f94SStefano Zampini } 4808e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4809e7931f94SStefano Zampini 4810e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4811e7931f94SStefano Zampini i = 1; 4812e7931f94SStefano Zampini if (color) i=0; 4813e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4814e7931f94SStefano Zampini /* get back IS */ 4815e7931f94SStefano Zampini *is_sends = ranks_send_to; 4816e7931f94SStefano Zampini PetscFunctionReturn(0); 4817e7931f94SStefano Zampini } 4818e7931f94SStefano Zampini 4819e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4820e7931f94SStefano Zampini 4821e7931f94SStefano Zampini #undef __FUNCT__ 4822e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 482353a05cb3SStefano 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[]) 4824e7931f94SStefano Zampini { 482570cf5478SStefano Zampini Mat local_mat; 4826e7931f94SStefano Zampini IS is_sends_internal; 48279d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 482828143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 48299d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4830e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4831e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4832e7931f94SStefano Zampini const PetscInt* is_indices; 4833e7931f94SStefano Zampini MatType new_local_type; 4834e7931f94SStefano Zampini /* buffers */ 4835e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 483628143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 48379d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4838e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4839e7931f94SStefano Zampini /* MPI */ 484028143c3dSStefano Zampini MPI_Comm comm,comm_n; 484128143c3dSStefano Zampini PetscSubcomm subcomm; 4842e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 484328143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 484428143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 484528143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 484628143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 484728143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4848e7931f94SStefano Zampini PetscErrorCode ierr; 4849e7931f94SStefano Zampini 4850e7931f94SStefano Zampini PetscFunctionBegin; 485128143c3dSStefano Zampini /* TODO: add missing checks */ 485228143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 485328143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 485428143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 485528143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4856e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 485728143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4858e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4859e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4860e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4861e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4862e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 486328143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 486470cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 486570cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 486628143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 486770cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 486870cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 486970cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 487070cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 487170cf5478SStefano Zampini } 4872e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4873e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4874e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4875e7931f94SStefano Zampini if (!is_sends) { 487628143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4877b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4878c8587f34SStefano Zampini } else { 4879e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4880e7931f94SStefano Zampini is_sends_internal = is_sends; 4881c8587f34SStefano Zampini } 4882e7931f94SStefano Zampini 4883e7931f94SStefano Zampini /* get comm */ 4884a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4885e7931f94SStefano Zampini 4886e7931f94SStefano Zampini /* compute number of sends */ 4887e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4888e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4889e7931f94SStefano Zampini 4890e7931f94SStefano Zampini /* compute number of receives */ 4891e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4892785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4893e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4894e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4895e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4896e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4897e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4898e7931f94SStefano Zampini 489928143c3dSStefano Zampini /* restrict comm if requested */ 490028143c3dSStefano Zampini subcomm = 0; 490128143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 490228143c3dSStefano Zampini if (restrict_comm) { 4903779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4904779c1cceSStefano Zampini 490528143c3dSStefano Zampini color = 0; 490653a05cb3SStefano Zampini if (restrict_full) { 490753a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 490853a05cb3SStefano Zampini } else { 490953a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 491053a05cb3SStefano Zampini } 491128143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 491228143c3dSStefano Zampini subcommsize = commsize - subcommsize; 491328143c3dSStefano Zampini /* check if reuse has been requested */ 491428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 491528143c3dSStefano Zampini if (*mat_n) { 491628143c3dSStefano Zampini PetscMPIInt subcommsize2; 491728143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 491828143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 491928143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 492028143c3dSStefano Zampini } else { 492128143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 492228143c3dSStefano Zampini } 492328143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4924779c1cceSStefano Zampini PetscMPIInt rank; 4925779c1cceSStefano Zampini 4926779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 492728143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 492828143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 492928143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4930306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 493128143c3dSStefano Zampini } 493228143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 493328143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 493428143c3dSStefano Zampini } else { 493528143c3dSStefano Zampini comm_n = comm; 493628143c3dSStefano Zampini } 493728143c3dSStefano Zampini 4938e7931f94SStefano Zampini /* prepare send/receive buffers */ 4939785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4940e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4941785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4942e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 494328143c3dSStefano Zampini if (nis) { 4944854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 494528143c3dSStefano Zampini } 4946e7931f94SStefano Zampini 494728143c3dSStefano Zampini /* Get data from local matrices */ 4948e7931f94SStefano Zampini if (!isdense) { 4949a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4950e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4951e7931f94SStefano Zampini /* 4952e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4953e7931f94SStefano Zampini send_buffer_idxs should contain: 4954e7931f94SStefano Zampini - MatType_PRIVATE type 4955e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4956e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4957e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4958e7931f94SStefano Zampini */ 4959e7931f94SStefano Zampini } else { 4960e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 49613bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4962854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4963e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4964e7931f94SStefano Zampini send_buffer_idxs[1] = i; 49653bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4966e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 49673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4968e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4969e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4970e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4971e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4972c8587f34SStefano Zampini } 4973c8587f34SStefano Zampini } 4974e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 497528143c3dSStefano Zampini /* additional is (if any) */ 497628143c3dSStefano Zampini if (nis) { 497728143c3dSStefano Zampini PetscMPIInt psum; 497828143c3dSStefano Zampini PetscInt j; 497928143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 498028143c3dSStefano Zampini PetscInt plen; 498128143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 498228143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 498328143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 498428143c3dSStefano Zampini } 4985854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 498628143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 498728143c3dSStefano Zampini PetscInt plen; 498828143c3dSStefano Zampini const PetscInt *is_array_idxs; 498928143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 499028143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 499128143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 499228143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 499328143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 499428143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 499528143c3dSStefano Zampini } 499628143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 499728143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 499828143c3dSStefano Zampini } 499928143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 500028143c3dSStefano Zampini } 500128143c3dSStefano Zampini 5002e7931f94SStefano Zampini buf_size_idxs = 0; 5003e7931f94SStefano Zampini buf_size_vals = 0; 500428143c3dSStefano Zampini buf_size_idxs_is = 0; 5005e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5006e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 5007e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 500828143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 5009e7931f94SStefano Zampini } 5010785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 5011785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 501295ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 5013e7931f94SStefano Zampini 5014e7931f94SStefano Zampini /* get new tags for clean communications */ 5015e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 5016e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 501728143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 5018e7931f94SStefano Zampini 5019e7931f94SStefano Zampini /* allocate for requests */ 5020785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 5021785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 502295ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 5023785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 5024785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 502595ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 5026e7931f94SStefano Zampini 5027e7931f94SStefano Zampini /* communications */ 5028e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5029e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 503028143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 5031e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5032e7931f94SStefano Zampini source_dest = onodes[i]; 5033e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 5034e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 5035e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5036e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 503728143c3dSStefano Zampini if (nis) { 503828143c3dSStefano 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); 503928143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 504028143c3dSStefano Zampini } 5041e7931f94SStefano Zampini } 5042e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5043e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5044e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5045e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 504628143c3dSStefano Zampini if (nis) { 504728143c3dSStefano 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); 504828143c3dSStefano Zampini } 5049e7931f94SStefano Zampini } 5050e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5051e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5052e7931f94SStefano Zampini 5053e7931f94SStefano Zampini /* assemble new l2g map */ 5054e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5055e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50569d30be91SStefano Zampini new_local_rows = 0; 5057e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50589d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5059e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5060e7931f94SStefano Zampini } 50619d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5062e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50639d30be91SStefano Zampini new_local_rows = 0; 5064e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50659d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 50669d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5067e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5068e7931f94SStefano Zampini } 50699d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 50709d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5071e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5072e7931f94SStefano Zampini 5073e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5074e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5075e7931f94SStefano 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) */ 5076e7931f94SStefano Zampini if (n_recvs) { 507728143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5078e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5079e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5080e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5081e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5082e7931f94SStefano Zampini break; 5083e7931f94SStefano Zampini } 5084e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5085e7931f94SStefano Zampini } 5086e7931f94SStefano Zampini switch (new_local_type_private) { 508728143c3dSStefano Zampini case MATDENSE_PRIVATE: 508828143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5089e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5090e7931f94SStefano Zampini bs = 1; 509128143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 509228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 509328143c3dSStefano Zampini bs = 1; 509428143c3dSStefano Zampini } 5095e7931f94SStefano Zampini break; 5096e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5097e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5098e7931f94SStefano Zampini bs = 1; 5099e7931f94SStefano Zampini break; 5100e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5101e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5102e7931f94SStefano Zampini break; 5103e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5104e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5105e7931f94SStefano Zampini break; 5106e7931f94SStefano Zampini default: 51079d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5108e7931f94SStefano Zampini break; 5109e7931f94SStefano Zampini } 511028143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 511128143c3dSStefano Zampini new_local_type = MATSEQDENSE; 511228143c3dSStefano Zampini bs = 1; 5113e7931f94SStefano Zampini } 5114e7931f94SStefano Zampini 511570cf5478SStefano Zampini /* create MATIS object if needed */ 511670cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5117e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5118e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 511970cf5478SStefano Zampini } else { 512070cf5478SStefano Zampini /* it also destroys the local matrices */ 512170cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 512270cf5478SStefano Zampini } 512370cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5124e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 51259d30be91SStefano Zampini 51269d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 51279d30be91SStefano Zampini 51289d30be91SStefano Zampini /* Global to local map of received indices */ 51299d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 51309d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 51319d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 51329d30be91SStefano Zampini 51339d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 51349d30be91SStefano Zampini buf_size_idxs = 0; 51359d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51369d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 51379d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 51389d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 51399d30be91SStefano Zampini } 51409d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 51419d30be91SStefano Zampini 51429d30be91SStefano Zampini /* set preallocation */ 51439d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 51449d30be91SStefano Zampini if (!newisdense) { 51459d30be91SStefano Zampini PetscInt *new_local_nnz=0; 51469d30be91SStefano Zampini 51479d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 51489d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 51499d30be91SStefano Zampini if (n_recvs) { 51509d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 51519d30be91SStefano Zampini } 51529d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51539d30be91SStefano Zampini PetscInt j; 51549d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 51559d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 51569d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 51579d30be91SStefano Zampini } 51589d30be91SStefano Zampini } else { 51599d30be91SStefano Zampini /* TODO */ 51609d30be91SStefano Zampini } 51619d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 51629d30be91SStefano Zampini } 51639d30be91SStefano Zampini if (new_local_nnz) { 51649d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 51659d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 51669d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 51679d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51689d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 51699d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51709d30be91SStefano Zampini } else { 51719d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51729d30be91SStefano Zampini } 51739d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 51749d30be91SStefano Zampini } else { 51759d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51769d30be91SStefano Zampini } 5177e7931f94SStefano Zampini 5178e7931f94SStefano Zampini /* set values */ 5179e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 51809d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5181e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5182e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5183e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 51849d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5185e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5186e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5187e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 518828143c3dSStefano Zampini } else { 518928143c3dSStefano Zampini /* TODO */ 5190e7931f94SStefano Zampini } 5191e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5192e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5193e7931f94SStefano Zampini } 5194e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5195e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 519670cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 519770cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 51989d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 51999d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5200e7931f94SStefano Zampini 5201dfd14d43SStefano Zampini #if 0 520228143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5203e7931f94SStefano Zampini Vec lvec,rvec; 5204e7931f94SStefano Zampini PetscReal infty_error; 5205e7931f94SStefano Zampini 52062a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5207e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5208e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5209e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 521070cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5211e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5212e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5213e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5214e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5215e7931f94SStefano Zampini } 521628143c3dSStefano Zampini #endif 5217e7931f94SStefano Zampini 521828143c3dSStefano Zampini /* assemble new additional is (if any) */ 521928143c3dSStefano Zampini if (nis) { 522028143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 522128143c3dSStefano Zampini 522228143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5223854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 522428143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 522528143c3dSStefano Zampini psum = 0; 522628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 522728143c3dSStefano Zampini for (j=0;j<nis;j++) { 522828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 522928143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 523028143c3dSStefano Zampini psum += plen; 523128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 523228143c3dSStefano Zampini } 523328143c3dSStefano Zampini } 5234854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5235854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 523628143c3dSStefano Zampini for (i=1;i<nis;i++) { 523728143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 523828143c3dSStefano Zampini } 523928143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 524028143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 524128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 524228143c3dSStefano Zampini for (j=0;j<nis;j++) { 524328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 524428143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 524528143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 524628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 524728143c3dSStefano Zampini } 524828143c3dSStefano Zampini } 524928143c3dSStefano Zampini for (i=0;i<nis;i++) { 525028143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 525128143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 525228143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 525328143c3dSStefano Zampini } 525428143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 525528143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 525628143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 525728143c3dSStefano Zampini } 5258e7931f94SStefano Zampini /* free workspace */ 525928143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5260e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5261e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5262e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5263e7931f94SStefano Zampini if (isdense) { 5264e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5265e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5266e7931f94SStefano Zampini } else { 5267e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5268e7931f94SStefano Zampini } 526928143c3dSStefano Zampini if (nis) { 527028143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 527128143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 527228143c3dSStefano Zampini } 5273e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5274e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 527528143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5276e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5277e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 527828143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5279e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5280e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5281e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5282e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5283e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 528428143c3dSStefano Zampini if (nis) { 528528143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 528628143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 528728143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 528828143c3dSStefano Zampini } 528928143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 529028143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 529128143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 529228143c3dSStefano Zampini for (i=0;i<nis;i++) { 529328143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 529428143c3dSStefano Zampini } 529553a05cb3SStefano Zampini *mat_n = NULL; 529628143c3dSStefano Zampini } 5297e7931f94SStefano Zampini PetscFunctionReturn(0); 5298e7931f94SStefano Zampini } 5299a57a6d2fSStefano Zampini 530012edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5301af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 530212edc857SStefano Zampini 5303c8587f34SStefano Zampini #undef __FUNCT__ 5304c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5305c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5306c8587f34SStefano Zampini { 5307c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5308c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 530920a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 53109881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 531120a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 53126e683305SStefano Zampini IS coarse_is,*isarray; 53136e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 531430368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5315f9eb5b7dSStefano Zampini PC pc_temp; 5316c8587f34SStefano Zampini PCType coarse_pc_type; 5317c8587f34SStefano Zampini KSPType coarse_ksp_type; 5318f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 53194f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 53206e683305SStefano Zampini Mat t_coarse_mat_is; 53216e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 53226e683305SStefano Zampini PetscMPIInt all_procs; 532374e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 532468457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 532522bc73bbSStefano Zampini PetscScalar *array; 53269881197aSStefano Zampini PetscErrorCode ierr; 5327fdc09c96SStefano Zampini 5328c8587f34SStefano Zampini PetscFunctionBegin; 5329c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 533068457ee5SStefano 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 */ 5331fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 53325a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5333fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5334f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5335f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5336f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5337fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 533851bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 533951bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5340dc4bcba2SStefano Zampini PC pc; 5341dc4bcba2SStefano Zampini PetscBool isbddc; 5342dc4bcba2SStefano Zampini 5343dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5344dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5345dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5346dc4bcba2SStefano Zampini if (isbddc) { 534763c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 534863c961adSStefano Zampini } else { 5349727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 535063c961adSStefano Zampini } 5351fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5352fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5353fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5354f4ddd8eeSStefano Zampini } 5355fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5356fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5357f4ddd8eeSStefano Zampini } 535870cf5478SStefano Zampini /* reset any subassembling information */ 535970cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 53606e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 53616e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5362fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5363f4ddd8eeSStefano Zampini } 5364c8587f34SStefano Zampini 53656e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 53662b510759SStefano Zampini im_active = !!(pcis->n); 53672b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 53686e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 53696e683305SStefano Zampini void_procs = all_procs-active_procs; 53706e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 537174e2c79eSStefano Zampini redist = PETSC_FALSE; 537222bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 53736e683305SStefano Zampini csin_ml = PETSC_TRUE; 53746e683305SStefano Zampini ncoarse_ml = void_procs; 5375779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5376779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 53776e683305SStefano Zampini csin_ds = PETSC_TRUE; 537818a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 537918a45a71SStefano Zampini redist = PETSC_TRUE; 538018a45a71SStefano Zampini } else { 53816e683305SStefano Zampini csin_ds = PETSC_TRUE; 5382779c1cceSStefano Zampini ncoarse_ds = active_procs; 5383779c1cceSStefano Zampini redist = PETSC_TRUE; 538418a45a71SStefano Zampini } 53856e683305SStefano Zampini } else { 53866e683305SStefano Zampini csin_ml = PETSC_FALSE; 53876e683305SStefano Zampini ncoarse_ml = all_procs; 53886e683305SStefano Zampini if (void_procs) { 53896e683305SStefano Zampini csin_ds = PETSC_TRUE; 53906e683305SStefano Zampini ncoarse_ds = void_procs; 53916e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 53926e683305SStefano Zampini } else { 5393779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 539474e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 539574e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 539674e2c79eSStefano Zampini redist = PETSC_TRUE; 539774e2c79eSStefano Zampini } else { 53986e683305SStefano Zampini csin_ds = PETSC_FALSE; 53996e683305SStefano Zampini ncoarse_ds = all_procs; 54006e683305SStefano Zampini } 54016e683305SStefano Zampini } 540274e2c79eSStefano Zampini } 54036e683305SStefano Zampini 54046e683305SStefano Zampini /* 54056e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 54066e683305SStefano Zampini - we have not exceeded the number of levels requested 54076e683305SStefano Zampini - we can actually subassemble the active processes 54086e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 54096e683305SStefano Zampini */ 54106e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 54116e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 54126e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 54136e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 54146e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5415f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 54162b510759SStefano Zampini } else { 5417f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5418c8587f34SStefano Zampini } 5419c8587f34SStefano Zampini } 54206e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 54216e683305SStefano Zampini if (multilevel_allowed) { 54226e683305SStefano Zampini ncoarse = ncoarse_ml; 54236e683305SStefano Zampini csin = csin_ml; 542458da7f69SStefano Zampini redist = PETSC_FALSE; 54256e683305SStefano Zampini } else { 54266e683305SStefano Zampini ncoarse = ncoarse_ds; 54276e683305SStefano Zampini csin = csin_ds; 54286e683305SStefano Zampini } 5429e7931f94SStefano Zampini 5430abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5431abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5432abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5433abbbba34SStefano Zampini 5434abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 543522bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 543622bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 543722bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 543822bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5439e176bc59SStefano 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); 54406e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 54416e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 54426e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5443abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5444abbbba34SStefano Zampini 54456e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 544630368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 54476e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 54486e683305SStefano Zampini const PetscInt *idxs; 54496e683305SStefano Zampini ISLocalToGlobalMapping tmap; 54506e683305SStefano Zampini 54516e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 54520be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 54536e683305SStefano Zampini /* allocate space for temporary storage */ 5454854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5455854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 54566e683305SStefano Zampini /* allocate for IS array */ 54576e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 54586e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 545930368db7SStefano Zampini nisvert = 0; 546030368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 546130368db7SStefano Zampini nisvert = 1; 546230368db7SStefano Zampini } 546330368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5464854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 54656e683305SStefano Zampini /* dofs splitting */ 54666e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 54676e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 54686e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 54696e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54706e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54716e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54726e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 547330368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 54746e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 54756e683305SStefano Zampini } 54766e683305SStefano Zampini /* neumann boundaries */ 54776e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 54786e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 54796e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 54806e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54816e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54826e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54836e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 548430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 54856e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 54866e683305SStefano Zampini } 548730368db7SStefano Zampini /* primal vertices (benign) */ 548830368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 548930368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 549030368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 549130368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 549230368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 549330368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 549430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 549530368db7SStefano Zampini } 54966e683305SStefano Zampini /* free memory */ 54976e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 54986e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 54996e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 55006e683305SStefano Zampini } else { 55016e683305SStefano Zampini nis = 0; 55026e683305SStefano Zampini nisdofs = 0; 55036e683305SStefano Zampini nisneu = 0; 550430368db7SStefano Zampini nisvert = 0; 55056e683305SStefano Zampini isarray = NULL; 55066e683305SStefano Zampini } 55076e683305SStefano Zampini /* destroy no longer needed map */ 55086e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 55096e683305SStefano Zampini 55106e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 55116e683305SStefano Zampini coarse_mat_is = NULL; 55126e683305SStefano Zampini if (csin) { 55136e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 551474e2c79eSStefano Zampini if (redist) { 551574e2c79eSStefano Zampini PetscMPIInt rank; 5516779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 551774e2c79eSStefano Zampini 551874e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 551958da7f69SStefano Zampini spc = active_procs/ncoarse; 552058da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5521779c1cceSStefano Zampini if (im_active) { 5522779c1cceSStefano Zampini destsize = 1; 552374e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 552474e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 552574e2c79eSStefano Zampini } else { 552674e2c79eSStefano Zampini dest[0] = rank/(spc+1); 552774e2c79eSStefano Zampini } 552874e2c79eSStefano Zampini } else { 5529779c1cceSStefano Zampini destsize = 0; 55306e683305SStefano Zampini } 5531779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5532779c1cceSStefano Zampini } else if (csin_type_simple) { 55336e683305SStefano Zampini PetscMPIInt rank; 55346e683305SStefano Zampini PetscInt issize,isidx; 5535779c1cceSStefano Zampini 55366e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 55376e683305SStefano Zampini if (im_active) { 55386e683305SStefano Zampini issize = 1; 55396e683305SStefano Zampini isidx = (PetscInt)rank; 55406e683305SStefano Zampini } else { 55416e683305SStefano Zampini issize = 0; 55426e683305SStefano Zampini isidx = -1; 55436e683305SStefano Zampini } 55446e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5545779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5546b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 55476e683305SStefano Zampini } 5548779c1cceSStefano Zampini 5549779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5550779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5551779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5552779c1cceSStefano Zampini PetscInt *coarse_candidates; 5553779c1cceSStefano Zampini const PetscInt* tisindices; 5554779c1cceSStefano Zampini 5555779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5556779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5557779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5558779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5559779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5560779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5561779c1cceSStefano Zampini } 5562779c1cceSStefano Zampini } 5563779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5564779c1cceSStefano Zampini 5565779c1cceSStefano Zampini 55666e683305SStefano Zampini if (pcbddc->dbg_flag) { 55676e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 55696e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55706e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5571779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 55726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 55736e683305SStefano Zampini } 55746e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 55756e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55766e683305SStefano Zampini } 55776e683305SStefano Zampini /* shift the pattern on coarse candidates */ 55786e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 55796e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5580854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 55816e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 55826e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 55836e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 55846e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 55856e683305SStefano Zampini } 55866e683305SStefano Zampini if (pcbddc->dbg_flag) { 55876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 55896e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55906e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55916e683305SStefano Zampini } 5592779c1cceSStefano Zampini } 55936e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 559453a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 559553a05cb3SStefano 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); 559653a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 559753a05cb3SStefano 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); 559853a05cb3SStefano Zampini } 55996e683305SStefano Zampini } else { 56006e683305SStefano Zampini if (pcbddc->dbg_flag) { 56016e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 56036e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56046e683305SStefano Zampini } 56056e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 56066e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 56076e683305SStefano Zampini } 56086e683305SStefano Zampini 56096e683305SStefano Zampini /* create local to global scatters for coarse problem */ 561068457ee5SStefano Zampini if (compute_vecs) { 56116e683305SStefano Zampini PetscInt lrows; 56126e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 56136e683305SStefano Zampini if (coarse_mat_is) { 56146e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 56156e683305SStefano Zampini } else { 56166e683305SStefano Zampini lrows = 0; 56176e683305SStefano Zampini } 56186e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 56196e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 56206e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 56216e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56226e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 56236e683305SStefano Zampini } 56246e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 56256e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5626c8587f34SStefano Zampini 5627f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5628f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5629f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5630f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5631f9eb5b7dSStefano Zampini } else { 5632f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5633f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5634c8587f34SStefano Zampini } 5635c8587f34SStefano Zampini 56366e683305SStefano Zampini /* print some info if requested */ 56376e683305SStefano Zampini if (pcbddc->dbg_flag) { 56386e683305SStefano Zampini if (!multilevel_allowed) { 56396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56406e683305SStefano Zampini if (multilevel_requested) { 56416e683305SStefano 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); 56426e683305SStefano Zampini } else if (pcbddc->max_levels) { 56436e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 56446e683305SStefano Zampini } 56456e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56466e683305SStefano Zampini } 56476e683305SStefano Zampini } 56486e683305SStefano Zampini 5649f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 56506e683305SStefano Zampini if (coarse_mat_is) { 56516e683305SStefano Zampini MatReuse coarse_mat_reuse; 56526a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 56536e683305SStefano Zampini if (pcbddc->dbg_flag) { 56546e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 56556e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56566e683305SStefano Zampini } 5657f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5658312be037SStefano Zampini char prefix[256],str_level[16]; 5659e604994aSStefano Zampini size_t len; 56606e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5661422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5662c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5663f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 56645f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5665c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 56666e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5667c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5668c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5669e604994aSStefano Zampini /* prefix */ 5670e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5671e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5672e604994aSStefano Zampini if (!pcbddc->current_level) { 5673e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5674e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5675c8587f34SStefano Zampini } else { 5676e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5677312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5678312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 567934d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5680312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5681e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5682e604994aSStefano Zampini } 5683e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 56843e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 56853e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 56863e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 56873e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5688f9eb5b7dSStefano Zampini /* allow user customization */ 5689f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 56903e3c6dadSStefano Zampini } 56913e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 569251bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 56933e3c6dadSStefano Zampini if (nisdofs) { 56943e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 56953e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 56963e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 56973e3c6dadSStefano Zampini } 56983e3c6dadSStefano Zampini } 56993e3c6dadSStefano Zampini if (nisneu) { 57003e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 57013e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5702312be037SStefano Zampini } 570330368db7SStefano Zampini if (nisvert) { 570430368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 570530368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 570630368db7SStefano Zampini } 5707f9eb5b7dSStefano Zampini 5708f9eb5b7dSStefano Zampini /* get some info after set from options */ 5709f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5710f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 57114f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 57126e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5713f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5714f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5715f9eb5b7dSStefano Zampini } 571639f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 57174f3a063dSStefano Zampini if (isredundant) { 57184f3a063dSStefano Zampini KSP inner_ksp; 57194f3a063dSStefano Zampini PC inner_pc; 57204f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 57214f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 57224f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 57234f3a063dSStefano Zampini } 5724f9eb5b7dSStefano Zampini 5725f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5726fa7f1dd8SStefano Zampini if (coarse_reuse) { 572781d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5728fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 57296e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5730fa7f1dd8SStefano Zampini } else { 57316e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5732fa7f1dd8SStefano Zampini } 5733c8587f34SStefano Zampini if (isbddc || isnn) { 5734720d30f9SStefano Zampini if (isbddc) { /* currently there are no APIs for these options */ 5735720d30f9SStefano Zampini PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data; 5736720d30f9SStefano Zampini pcbddc_coarse->detect_disconnected = PETSC_TRUE; 5737720d30f9SStefano Zampini pcbddc_coarse->benign_saddle_point = pcbddc->benign_saddle_point; 5738720d30f9SStefano Zampini pcbddc_coarse->benign_compute_nonetflux = pcbddc->benign_compute_nonetflux; 5739720d30f9SStefano Zampini if (pcbddc_coarse->benign_compute_nonetflux) { 5740720d30f9SStefano Zampini pcbddc_coarse->adaptive_userdefined = PETSC_TRUE; 5741720d30f9SStefano Zampini } 5742d4d8cf7bSStefano Zampini } 574322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 574470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5745b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 574622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 57476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 57496e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 57506e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 575122b6e8a2SStefano Zampini } 575270cf5478SStefano Zampini } 575353a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 575470cf5478SStefano Zampini } else { 575522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 575622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 575722bc73bbSStefano Zampini } 575822bc73bbSStefano Zampini } else { 57592e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5760c8587f34SStefano Zampini } 5761c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5762c8587f34SStefano Zampini 57633301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 57645a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 57653301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 57663301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 57673301b35fSStefano Zampini } 57683301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 57693301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 57703301b35fSStefano Zampini } 57713301b35fSStefano Zampini if (pc->pmat->spd_set) { 57723301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 57733301b35fSStefano Zampini } 57746e683305SStefano Zampini /* set operators */ 57755f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 57766e683305SStefano Zampini if (pcbddc->dbg_flag) { 57776e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57786e683305SStefano Zampini } 57796e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 57806e683305SStefano Zampini coarse_mat = 0; 57816e683305SStefano Zampini } 57826e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5783b1ecc7b1SStefano Zampini #if 0 5784b9b85e73SStefano Zampini { 5785b9b85e73SStefano Zampini PetscViewer viewer; 5786b9b85e73SStefano Zampini char filename[256]; 5787b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5788b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5789b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5790b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5791b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5792b9b85e73SStefano Zampini } 5793b9b85e73SStefano Zampini #endif 5794c8587f34SStefano Zampini 5795c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5796298c0119SStefano Zampini #if 0 5797c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5798c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 579998a51de6SStefano Zampini } 5800298c0119SStefano Zampini #endif 5801b0f5fe93SStefano Zampini /* hack */ 580298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 580398a51de6SStefano Zampini Vec crhs,csol; 580404708bb6SStefano Zampini 5805f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5806f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5807f347579bSStefano Zampini if (!csol) { 58082a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5809f9eb5b7dSStefano Zampini } 5810f347579bSStefano Zampini if (!crhs) { 58112a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5812f347579bSStefano Zampini } 5813b0f5fe93SStefano Zampini } 5814b0f5fe93SStefano Zampini 5815b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5816b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5817b0f5fe93SStefano Zampini 5818b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 58194f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 58204f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 58214f1b2e48SStefano Zampini } 5822b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5823b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5824b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5825b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5826b0f5fe93SStefano Zampini if (coarse_mat) { 5827b0f5fe93SStefano Zampini Vec nullv; 5828b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5829b0f5fe93SStefano Zampini PetscInt nl; 5830b0f5fe93SStefano Zampini 5831b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5832b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5833b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5834b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5835b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5836b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5837b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5838b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5839b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5840b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5841b0f5fe93SStefano Zampini } 5842b0f5fe93SStefano Zampini } 5843b0f5fe93SStefano Zampini 5844b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5845b0f5fe93SStefano Zampini PetscBool ispreonly; 5846b0f5fe93SStefano Zampini 5847b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5848b0f5fe93SStefano Zampini PetscBool isnull; 5849b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5850d4d8cf7bSStefano Zampini if (0) { 585130368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5852b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5853b0f5fe93SStefano Zampini } else { 5854b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5855b0f5fe93SStefano Zampini } 5856b0f5fe93SStefano Zampini } else { 5857b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5858b0f5fe93SStefano Zampini } 5859b0f5fe93SStefano Zampini } 5860b0f5fe93SStefano Zampini /* setup coarse ksp */ 5861b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5862cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5863cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 58646e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5865c8587f34SStefano Zampini KSP check_ksp; 58662b510759SStefano Zampini KSPType check_ksp_type; 5867c8587f34SStefano Zampini PC check_pc; 58686e683305SStefano Zampini Vec check_vec,coarse_vec; 58696a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 58702b510759SStefano Zampini PetscInt its; 58716e683305SStefano Zampini PetscBool compute_eigs; 58726e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 58736e683305SStefano Zampini PetscInt neigs; 58748e185a42SStefano Zampini const char *prefix; 5875c8587f34SStefano Zampini 58762b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 58776e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5878422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 587923ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5880f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 58812b510759SStefano Zampini if (ispreonly) { 58822b510759SStefano Zampini check_ksp_type = KSPPREONLY; 58836e683305SStefano Zampini compute_eigs = PETSC_FALSE; 58842b510759SStefano Zampini } else { 5885cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 58866e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5887c8587f34SStefano Zampini } 5888c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 58896e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 58906e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 58916e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5892a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5893a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5894a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5895a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5896c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5897c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5898c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5899c8587f34SStefano Zampini /* create random vec */ 59006e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 59016e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5902c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5903c8587f34SStefano Zampini if (CoarseNullSpace) { 5904c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5905c8587f34SStefano Zampini } 59066e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5907c8587f34SStefano Zampini /* solve coarse problem */ 59086e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5909c8587f34SStefano Zampini if (CoarseNullSpace) { 59106e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5911c8587f34SStefano Zampini } 5912cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 59136e683305SStefano Zampini if (compute_eigs) { 5914854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5915854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 59166e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 59176e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 59186e683305SStefano Zampini lambda_min = eigs_r[0]; 59196e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 59206e683305SStefano Zampini if (lambda_max>lambda_min) { 5921cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5922cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5923cbcc2c2aSStefano Zampini } 5924c8587f34SStefano Zampini } 5925c8587f34SStefano Zampini } 5926cbcc2c2aSStefano Zampini 5927c8587f34SStefano Zampini /* check coarse problem residual error */ 59286e683305SStefano Zampini if (pcbddc->dbg_flag) { 59296e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 59306e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59316e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5932c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 59336e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 59346e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5935c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5936779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 59376e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 59386e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 59396e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 59406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5941b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5942b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5943b0f5fe93SStefano Zampini } 59446e683305SStefano Zampini if (compute_eigs) { 59456e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5946deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5947c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 59486e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 59496e683305SStefano 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); 59506e683305SStefano Zampini for (i=0;i<neigs;i++) { 59516e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5952c8587f34SStefano Zampini } 59536e683305SStefano Zampini } 59546e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 59556e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59566e683305SStefano Zampini } 5957c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 59586e683305SStefano Zampini if (compute_eigs) { 59596e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 59606e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5961c8587f34SStefano Zampini } 59626e683305SStefano Zampini } 59636e683305SStefano Zampini } 5964cbcc2c2aSStefano Zampini /* print additional info */ 5965cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 59666e683305SStefano Zampini /* waits until all processes reaches this point */ 59676e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5968cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5969cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5970cbcc2c2aSStefano Zampini } 5971cbcc2c2aSStefano Zampini 59722b510759SStefano Zampini /* free memory */ 5973c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5974fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5975c8587f34SStefano Zampini PetscFunctionReturn(0); 5976c8587f34SStefano Zampini } 5977674ae819SStefano Zampini 5978f34684f1SStefano Zampini #undef __FUNCT__ 5979f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5980f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5981f34684f1SStefano Zampini { 5982f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5983f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5984f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5985dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5986dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 598773be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5988dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5989f34684f1SStefano Zampini PetscErrorCode ierr; 5990f34684f1SStefano Zampini 5991f34684f1SStefano Zampini PetscFunctionBegin; 5992f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 59930e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 59940e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5995727cdba6SStefano Zampini } 5996dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 59973bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5998dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5999dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 6000dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 6001dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 6002dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 6003dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 60040e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 60050e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 60060e6343abSStefano Zampini } 6007dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 6008dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6009dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 6010dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 6011dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 6012f34684f1SStefano Zampini 6013f34684f1SStefano Zampini /* check numbering */ 6014f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 6015019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 6016dc456d91SStefano Zampini PetscInt i; 6017b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 6018f34684f1SStefano Zampini 6019f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6020f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 6021f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 60221575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6023019a44ceSStefano Zampini /* counter */ 6024019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6025019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 6026019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6027019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6028019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6029019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6030f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 6031f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 6032727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 6033f34684f1SStefano Zampini } 6034f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6035f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6036f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6037e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6038e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6039e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6040e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6041f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6042019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6043f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6044019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 604575c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 604675c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6047b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6048019a44ceSStefano 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); 6049f34684f1SStefano Zampini } 6050f34684f1SStefano Zampini } 6051019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6052b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6053f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6054f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6055f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6056f34684f1SStefano Zampini } 6057f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6058f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6059e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6060e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6061f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6062f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6063b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6064ca8b9ea9SStefano Zampini PetscInt *gidxs; 6065ca8b9ea9SStefano Zampini 6066ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 60673bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6068f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6069f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6070f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6071f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 60724bc2dc4bSStefano 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); 6073f34684f1SStefano Zampini } 6074f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6075ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6076f34684f1SStefano Zampini } 6077f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 60781575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6079302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6080f34684f1SStefano Zampini } 60818bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6082f34684f1SStefano Zampini /* get back data */ 6083f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6084f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6085674ae819SStefano Zampini PetscFunctionReturn(0); 6086674ae819SStefano Zampini } 6087674ae819SStefano Zampini 6088e456f2a8SStefano Zampini #undef __FUNCT__ 6089e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6090a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6091e456f2a8SStefano Zampini { 6092e456f2a8SStefano Zampini IS localis_t; 6093a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6094e456f2a8SStefano Zampini PetscScalar *vals; 6095e456f2a8SStefano Zampini PetscErrorCode ierr; 6096e456f2a8SStefano Zampini 6097e456f2a8SStefano Zampini PetscFunctionBegin; 6098a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6099e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6100854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6101e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6102e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6103a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6104a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 61051035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6106a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 61071035eff8SStefano Zampini } 6108a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6109e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6110e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6111a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6112a7dc3881SStefano Zampini /* now compute set in local ordering */ 6113a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6114a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6115a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6116a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6117a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6118ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6119e456f2a8SStefano Zampini lsize++; 6120e456f2a8SStefano Zampini } 6121e456f2a8SStefano Zampini } 6122854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6123a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6124ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6125e456f2a8SStefano Zampini idxs[lsize++] = i; 6126e456f2a8SStefano Zampini } 6127e456f2a8SStefano Zampini } 6128a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6129a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6130e456f2a8SStefano Zampini *localis = localis_t; 6131e456f2a8SStefano Zampini PetscFunctionReturn(0); 6132e456f2a8SStefano Zampini } 6133906d46d4SStefano Zampini 6134906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6135906d46d4SStefano Zampini #undef __FUNCT__ 6136906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6137906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6138906d46d4SStefano Zampini { 6139906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6140906d46d4SStefano Zampini PetscErrorCode ierr; 6141906d46d4SStefano Zampini 6142906d46d4SStefano Zampini PetscFunctionBegin; 6143906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6144906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6145906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6146906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6147906d46d4SStefano Zampini PetscFunctionReturn(0); 6148906d46d4SStefano Zampini } 6149906d46d4SStefano Zampini 6150906d46d4SStefano Zampini #undef __FUNCT__ 6151906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6152906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6153906d46d4SStefano Zampini { 6154906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6155906d46d4SStefano Zampini PetscErrorCode ierr; 6156906d46d4SStefano Zampini 6157906d46d4SStefano Zampini PetscFunctionBegin; 6158906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6159906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6160906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6161906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6162906d46d4SStefano Zampini PetscFunctionReturn(0); 6163906d46d4SStefano Zampini } 6164b96c3477SStefano Zampini 6165b96c3477SStefano Zampini #undef __FUNCT__ 6166b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 616708122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6168b96c3477SStefano Zampini { 6169a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6170b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6171b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6172a64f4aa4SStefano Zampini Mat S_j; 6173b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6174b96c3477SStefano Zampini PetscBool free_used_adj; 6175b96c3477SStefano Zampini PetscErrorCode ierr; 6176b96c3477SStefano Zampini 6177b96c3477SStefano Zampini PetscFunctionBegin; 6178b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6179b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 618008122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6181b96c3477SStefano Zampini used_xadj = NULL; 6182b96c3477SStefano Zampini used_adjncy = NULL; 6183b96c3477SStefano Zampini } else { 618408122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 618508122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 618608122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 618708122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6188b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6189b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6190b96c3477SStefano Zampini } else { 61912fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6192b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6193b96c3477SStefano Zampini PetscInt nvtxs; 6194b96c3477SStefano Zampini 61952fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 61962fffb893SStefano Zampini if (flg_row) { 6197b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6198b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6199b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6200b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 62012fffb893SStefano Zampini } else { 62022fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 62032fffb893SStefano Zampini used_xadj = NULL; 62042fffb893SStefano Zampini used_adjncy = NULL; 62052fffb893SStefano Zampini } 62062fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6207b96c3477SStefano Zampini } 6208b96c3477SStefano Zampini } 6209d5574798SStefano Zampini 6210d5574798SStefano Zampini /* setup sub_schurs data */ 6211a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6212df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6213df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6214a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 621572b8c272SStefano 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,NULL,NULL);CHKERRQ(ierr); 6216a64f4aa4SStefano Zampini } else { 62176816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 621804708bb6SStefano Zampini PetscBool isseqaij; 6219a3df083aSStefano Zampini PetscInt benign_n; 622072b8c272SStefano Zampini Mat change = NULL; 622172b8c272SStefano Zampini IS change_primal = NULL; 6222a3df083aSStefano Zampini 62235feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 62245feab87aSStefano Zampini PetscInt n_vertices; 62255feab87aSStefano Zampini 62265feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 62272034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 62285feab87aSStefano Zampini } 622904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 623004708bb6SStefano Zampini if (!isseqaij) { 623104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 623204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 623304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 623404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 623504708bb6SStefano Zampini } else { 623604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 623704708bb6SStefano Zampini } 623804708bb6SStefano Zampini } 6239a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6240a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6241ca92afb2SStefano Zampini } else { 6242a3df083aSStefano Zampini benign_n = 0; 6243ca92afb2SStefano Zampini } 6244bf3a8328SStefano Zampini /* If the user defines additional constraints, we import them here. 6245bf3a8328SStefano Zampini We need to compute the change of basis according to the quadrature weights attached to pmat via MatSetNearNullSpace, and this could not be done, at the moment, without some hacking */ 624672b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 624788c03ad3SStefano Zampini PC_IS *pcisf; 624888c03ad3SStefano Zampini PC_BDDC *pcbddcf; 624988c03ad3SStefano Zampini PC pcf; 625088c03ad3SStefano Zampini 625188c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 625288c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 625388c03ad3SStefano Zampini } 625488c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 625588c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 625688c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 625788c03ad3SStefano Zampini /* hacks */ 625888c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 625972b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 626072b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 626172b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 626272b8c272SStefano Zampini pcisf->n = pcis->n; 626372b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 626488c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 626588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 626688c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 626788c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 626888c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 626988c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 627072b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 627188c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 627288c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 627372b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 627472b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 627572b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 627672b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 627772b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 627888c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 627972b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 628088c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 628188c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 628288c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 628388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 628488c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 628588c03ad3SStefano Zampini pcf->ops->destroy = NULL; 628688c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 628788c03ad3SStefano Zampini } 628872b8c272SStefano 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,change,change_primal);CHKERRQ(ierr); 628972b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 629072b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6291ca92afb2SStefano Zampini } 6292d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6293b96c3477SStefano Zampini 6294b96c3477SStefano Zampini /* free adjacency */ 6295b96c3477SStefano Zampini if (free_used_adj) { 6296b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6297b96c3477SStefano Zampini } 6298b96c3477SStefano Zampini PetscFunctionReturn(0); 6299b96c3477SStefano Zampini } 6300b96c3477SStefano Zampini 6301b96c3477SStefano Zampini #undef __FUNCT__ 6302b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 630308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6304b96c3477SStefano Zampini { 6305b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6306b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6307b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6308b96c3477SStefano Zampini PCBDDCGraph graph; 6309b96c3477SStefano Zampini PetscErrorCode ierr; 6310b96c3477SStefano Zampini 6311b96c3477SStefano Zampini PetscFunctionBegin; 6312b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 631308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 63143301b35fSStefano Zampini IS verticesIS,verticescomm; 63153301b35fSStefano Zampini PetscInt vsize,*idxs; 6316b96c3477SStefano Zampini 6317b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 63183301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 63193301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63203301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 63213301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 63223301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6323b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 63247fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 63253301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 63263301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6327b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6328b96c3477SStefano Zampini /* 6329b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6330b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6331b96c3477SStefano Zampini } 6332b96c3477SStefano Zampini */ 6333b96c3477SStefano Zampini } else { 6334b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6335b96c3477SStefano Zampini } 6336b96c3477SStefano Zampini 6337b96c3477SStefano Zampini /* sub_schurs init */ 6338a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6339a64f4aa4SStefano Zampini 6340b96c3477SStefano Zampini /* free graph struct */ 634108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6342b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6343b96c3477SStefano Zampini } 6344b96c3477SStefano Zampini PetscFunctionReturn(0); 6345b96c3477SStefano Zampini } 6346fa34dd3eSStefano Zampini 6347fa34dd3eSStefano Zampini #undef __FUNCT__ 6348fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6349fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6350fa34dd3eSStefano Zampini { 6351fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6352fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6353fa34dd3eSStefano Zampini PetscErrorCode ierr; 6354fa34dd3eSStefano Zampini 6355fa34dd3eSStefano Zampini PetscFunctionBegin; 6356fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6357fa34dd3eSStefano Zampini IS zerodiag = NULL; 63584f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6359fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 63604f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 636175c01103SStefano Zampini PetscReal norm; 6362fa34dd3eSStefano Zampini PetscInt i; 6363fa34dd3eSStefano Zampini 6364fa34dd3eSStefano Zampini /* B0 and B0_B */ 6365fa34dd3eSStefano Zampini if (zerodiag) { 6366fa34dd3eSStefano Zampini IS dummy; 6367fa34dd3eSStefano Zampini 63684f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 63694f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6370fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6371fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6372fa34dd3eSStefano Zampini } 6373fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6374fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6375fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6376fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6377fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6378fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6379fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6380fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6381fa34dd3eSStefano Zampini /* S_j */ 6382fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6383fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6384fa34dd3eSStefano Zampini 6385fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6386fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6387fa34dd3eSStefano Zampini /* continuous in primal space */ 6388fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6389fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6390fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6391fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 63924f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 63934f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6394fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6395fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6396fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6397fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6398fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6399fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6400fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6401fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6402fa34dd3eSStefano Zampini 6403fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6404fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6405fa34dd3eSStefano Zampini /* local with Schur */ 6406fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6407fa34dd3eSStefano Zampini if (zerodiag) { 6408fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64094f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6410fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6411fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6412fa34dd3eSStefano Zampini } 6413fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6414fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6415fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6416fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6417fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6418fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6419fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6420fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6421fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6422fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6423fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6424fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6425fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6426fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6427fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6428fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6429fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6430fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6431fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6432fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6433fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6434fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6435fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6436fa34dd3eSStefano Zampini if (zerodiag) { 6437fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6438fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 64394f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6440fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6441fa34dd3eSStefano Zampini } 6442fa34dd3eSStefano Zampini /* BDDC */ 6443fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6444fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6445fa34dd3eSStefano Zampini 6446fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6447fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6448fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6449fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64504f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64514f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6452fa34dd3eSStefano Zampini } 64534f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6454fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6455fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6456fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6457fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6458fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6459fa34dd3eSStefano Zampini } 6460fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6461fa34dd3eSStefano Zampini } 6462