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 6494f1b2e48SStefano Zampini if (has_null_pressures) { 6504f1b2e48SStefano Zampini IS zerodiagc; 6514f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 6524f1b2e48SStefano Zampini PetscInt i,s,*nnz; 6534f1b2e48SStefano Zampini 6544f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 655339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 656339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 657339f8db1SStefano Zampini /* local change of basis for pressures */ 658339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 65997d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 660339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 661339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 662339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 6634f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 6644f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 6654f1b2e48SStefano Zampini PetscInt nzs,j; 6664f1b2e48SStefano Zampini 6674f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 6684f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6694f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 6704f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 6714f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 6724f1b2e48SStefano Zampini } 673339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 674339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 675339f8db1SStefano Zampini /* set identity on velocities */ 676339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 677339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 678339f8db1SStefano Zampini } 6794f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 6804f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 6819f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 6824f1b2e48SStefano 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); 683339f8db1SStefano Zampini /* set change on pressures */ 6844f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 6854f1b2e48SStefano Zampini PetscScalar *array; 6864f1b2e48SStefano Zampini PetscInt nzs; 6874f1b2e48SStefano Zampini 6884f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 6894f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 6904f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 691339f8db1SStefano Zampini PetscScalar vals[2]; 692339f8db1SStefano Zampini PetscInt cols[2]; 693339f8db1SStefano Zampini 694339f8db1SStefano Zampini cols[0] = idxs[i]; 6954f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 696339f8db1SStefano Zampini vals[0] = 1.; 697b0f5fe93SStefano Zampini vals[1] = 1.; 6984f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 699339f8db1SStefano Zampini } 7004f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7014f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7024f1b2e48SStefano Zampini array[nzs-1] = 1.; 7034f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7044f1b2e48SStefano Zampini /* store local idxs for p0 */ 7054f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7064f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 707339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7084f1b2e48SStefano Zampini } 709339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 710339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 711a3df083aSStefano Zampini /* project if needed */ 712a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 71397d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 714339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 71597d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 71697d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 717a3df083aSStefano Zampini } 7184f1b2e48SStefano Zampini /* store global idxs for p0 */ 7194f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 720339f8db1SStefano Zampini } 721ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7224f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 723b0f5fe93SStefano Zampini 724b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 725b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 726339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 727339f8db1SStefano Zampini PetscFunctionReturn(0); 728339f8db1SStefano Zampini } 729339f8db1SStefano Zampini 730339f8db1SStefano Zampini #undef __FUNCT__ 731015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 732015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 733efc2fbd9SStefano Zampini { 734efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 735efc2fbd9SStefano Zampini PetscErrorCode ierr; 736efc2fbd9SStefano Zampini 737efc2fbd9SStefano Zampini PetscFunctionBegin; 738efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 739efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 7404f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 741efc2fbd9SStefano Zampini } 742015636ebSStefano Zampini if (get) { /* use SF to get values */ 743efc2fbd9SStefano Zampini PetscScalar *array; 744efc2fbd9SStefano Zampini 745efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7464f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 7474f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 748efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 7494f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 7504f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 751efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 752efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 753efc2fbd9SStefano Zampini } 754efc2fbd9SStefano Zampini PetscFunctionReturn(0); 755efc2fbd9SStefano Zampini } 756efc2fbd9SStefano Zampini 757efc2fbd9SStefano Zampini #undef __FUNCT__ 758c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 759c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 760c263805aSStefano Zampini { 761c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 762c263805aSStefano Zampini PetscErrorCode ierr; 763c263805aSStefano Zampini 764c263805aSStefano Zampini PetscFunctionBegin; 765c263805aSStefano Zampini /* TODO: add error checking 766c263805aSStefano Zampini - avoid nested pop (or push) calls. 767c263805aSStefano Zampini - cannot push before pop. 7681c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 769c263805aSStefano Zampini */ 7704f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 771efc2fbd9SStefano Zampini PetscFunctionReturn(0); 772efc2fbd9SStefano Zampini } 773c263805aSStefano Zampini if (pop) { 774a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 7754f1b2e48SStefano Zampini IS is_p0; 7764f1b2e48SStefano Zampini MatReuse reuse; 777c263805aSStefano Zampini 778c263805aSStefano Zampini /* extract B_0 */ 7794f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 7804f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 7814f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 7824f1b2e48SStefano Zampini } 7834f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 7844f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 785c263805aSStefano Zampini /* remove rows and cols from local problem */ 786c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 78797d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 7884f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 7894f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 790a3df083aSStefano Zampini } else { 791a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 792a3df083aSStefano Zampini PetscScalar *vals; 793a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 794a3df083aSStefano Zampini 795a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 796a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 797a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 7980b5adadeSStefano Zampini PetscInt *nnz; 799a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 800a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 801a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 802331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 803331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 804331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 805331e053bSStefano Zampini nnz[i] = n - nnz[i]; 806331e053bSStefano Zampini } 807331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 808331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 809331e053bSStefano Zampini } 810a3df083aSStefano Zampini 811a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 812a3df083aSStefano Zampini PetscScalar *array; 813a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 814a3df083aSStefano Zampini 815a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 816a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 817a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 818a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 819a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 820a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 821a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 822a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 823a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 824a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 825a3df083aSStefano Zampini cum = 0; 826a3df083aSStefano Zampini for (j=0;j<n;j++) { 827a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 828a3df083aSStefano Zampini vals[cum] = array[j]; 829a3df083aSStefano Zampini idxs_ins[cum] = j; 830a3df083aSStefano Zampini cum++; 831a3df083aSStefano Zampini } 832a3df083aSStefano Zampini } 833a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 834a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 835a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 836a3df083aSStefano Zampini } 837a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 838a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 839a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 840a3df083aSStefano Zampini } 841c263805aSStefano Zampini } else { /* push */ 842a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8434f1b2e48SStefano Zampini PetscInt i; 8444f1b2e48SStefano Zampini 8454f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 8464f1b2e48SStefano Zampini PetscScalar *B0_vals; 8474f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 8484f1b2e48SStefano Zampini 8494f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8504f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8517b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 8524f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 8534f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 8544f1b2e48SStefano Zampini } 855c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 856c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 857a3df083aSStefano Zampini } else { 858a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 859a3df083aSStefano Zampini } 860c263805aSStefano Zampini } 861c263805aSStefano Zampini PetscFunctionReturn(0); 862c263805aSStefano Zampini } 863c263805aSStefano Zampini 864c263805aSStefano Zampini #undef __FUNCT__ 865b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 86608122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 867b1b3d7a2SStefano Zampini { 868b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 86908122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 87008122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 87108122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 87208122e43SStefano Zampini PetscScalar *work,lwork; 87308122e43SStefano Zampini PetscScalar *St,*S,*eigv; 87408122e43SStefano Zampini PetscScalar *Sarray,*Starray; 87508122e43SStefano Zampini PetscReal *eigs,thresh; 8761b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 877f6f667cfSStefano Zampini PetscBool allocated_S_St; 87808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 87908122e43SStefano Zampini PetscReal *rwork; 88008122e43SStefano Zampini #endif 881b1b3d7a2SStefano Zampini PetscErrorCode ierr; 882b1b3d7a2SStefano Zampini 883b1b3d7a2SStefano Zampini PetscFunctionBegin; 884df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 885df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 88608122e43SStefano Zampini } 88708122e43SStefano Zampini 88806a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 88906a4e24aSStefano 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); 89006a4e24aSStefano Zampini } 89106a4e24aSStefano Zampini 892fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 893fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 894fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 895fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 8961575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 897fd14bc51SStefano Zampini } 898fd14bc51SStefano Zampini 899e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 900e496cd5dSStefano 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); 901e496cd5dSStefano Zampini } 902e496cd5dSStefano Zampini 90308122e43SStefano Zampini /* max size of subsets */ 90408122e43SStefano Zampini mss = 0; 90508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 90608122e43SStefano Zampini PetscInt subset_size; 907862806e4SStefano Zampini 90808122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 90908122e43SStefano Zampini mss = PetscMax(mss,subset_size); 91008122e43SStefano Zampini } 91108122e43SStefano Zampini 91208122e43SStefano Zampini /* min/max and threshold */ 91308122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 914f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 91508122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 916f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 917f6f667cfSStefano Zampini if (nmin) { 918f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 919f6f667cfSStefano Zampini } 92008122e43SStefano Zampini 92108122e43SStefano Zampini /* allocate lapack workspace */ 92208122e43SStefano Zampini cum = cum2 = 0; 92308122e43SStefano Zampini maxneigs = 0; 92408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 92508122e43SStefano Zampini PetscInt n,subset_size; 926f6f667cfSStefano Zampini 92708122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 92808122e43SStefano Zampini n = PetscMin(subset_size,nmax); 9299162d606SStefano Zampini cum += subset_size; 9309162d606SStefano Zampini cum2 += subset_size*n; 93108122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 93208122e43SStefano Zampini } 93308122e43SStefano Zampini if (mss) { 9349ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 93508122e43SStefano Zampini PetscBLASInt B_itype = 1; 93608122e43SStefano Zampini PetscBLASInt B_N = mss; 9374c6709b3SStefano Zampini PetscReal zero = 0.0; 9384c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 93908122e43SStefano Zampini 94008122e43SStefano Zampini B_lwork = -1; 94108122e43SStefano Zampini S = NULL; 94208122e43SStefano Zampini St = NULL; 943a58a30b4SStefano Zampini eigs = NULL; 944a58a30b4SStefano Zampini eigv = NULL; 945a58a30b4SStefano Zampini B_iwork = NULL; 946a58a30b4SStefano Zampini B_ifail = NULL; 947d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 948d1710679SStefano Zampini rwork = NULL; 949d1710679SStefano Zampini #endif 9508bec7fa6SStefano Zampini thresh = 1.0; 95108122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 95208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 95308122e43SStefano 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)); 95408122e43SStefano Zampini #else 95508122e43SStefano 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)); 95608122e43SStefano Zampini #endif 95708122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 95808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 95908122e43SStefano Zampini } else { 96008122e43SStefano Zampini /* TODO */ 96108122e43SStefano Zampini } 96208122e43SStefano Zampini } else { 96308122e43SStefano Zampini lwork = 0; 96408122e43SStefano Zampini } 96508122e43SStefano Zampini 96608122e43SStefano Zampini nv = 0; 967d62866d3SStefano 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) */ 968d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 96908122e43SStefano Zampini } 9704c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 971f6f667cfSStefano Zampini if (allocated_S_St) { 972f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 973f6f667cfSStefano Zampini } 974f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 97508122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 97608122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 97708122e43SStefano Zampini #endif 9789162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 9799162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 9809162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 98108122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 9829162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 98308122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 98408122e43SStefano Zampini 98508122e43SStefano Zampini maxneigs = 0; 98608122e43SStefano Zampini cum = cum2 = cumarray = 0; 9879162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 9889162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 989d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 99008122e43SStefano Zampini const PetscInt *idxs; 99108122e43SStefano Zampini 992d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 99308122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 99408122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 99508122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 99608122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 9979162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 9989162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 99908122e43SStefano Zampini } 100008122e43SStefano Zampini cum2 = cum; 1001d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 100208122e43SStefano Zampini } 100308122e43SStefano Zampini 100408122e43SStefano Zampini if (mss) { /* multilevel */ 100508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 100608122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 100708122e43SStefano Zampini } 100808122e43SStefano Zampini 1009ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 101008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 101108122e43SStefano Zampini const PetscInt *idxs; 1012f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1013862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 101408122e43SStefano Zampini PetscBLASInt B_N; 1015aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 101608122e43SStefano Zampini 1017862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1018ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1019f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1020f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10219ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1022aff50787SStefano Zampini PetscInt j,k; 1023aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1024aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1025aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 102608122e43SStefano Zampini } 102708122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1028aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1029aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1030aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1031aff50787SStefano Zampini } 103208122e43SStefano Zampini } 103308122e43SStefano Zampini } else { 103408122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 103508122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 103608122e43SStefano Zampini } 10378bec7fa6SStefano Zampini } else { 1038f6f667cfSStefano Zampini S = Sarray + cumarray; 1039f6f667cfSStefano Zampini St = Starray + cumarray; 10408bec7fa6SStefano Zampini } 104108122e43SStefano Zampini 1042aff50787SStefano Zampini /* see if we can save some work */ 1043aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1044aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1045aff50787SStefano Zampini } 1046aff50787SStefano Zampini 1047aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1048aff50787SStefano Zampini B_neigs = 0; 1049aff50787SStefano Zampini } else { 10509ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 105108122e43SStefano Zampini PetscBLASInt B_itype = 1; 1052f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 10534c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 10549552c7c7SStefano Zampini PetscInt nmin_s; 105508122e43SStefano Zampini 1056fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 10578bec7fa6SStefano 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]]); 1058fd14bc51SStefano Zampini } 1059d16cbb6bSStefano Zampini 106008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1061d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1062d16cbb6bSStefano Zampini 1063d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 106408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1065f6f667cfSStefano 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)); 106608122e43SStefano Zampini #else 1067f6f667cfSStefano 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)); 106808122e43SStefano Zampini #endif 1069d16cbb6bSStefano Zampini } else { 1070d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1071d16cbb6bSStefano Zampini B_IL = 1; 1072d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1073d16cbb6bSStefano 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)); 1074d16cbb6bSStefano Zampini #else 1075d16cbb6bSStefano 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)); 1076d16cbb6bSStefano Zampini #endif 1077d16cbb6bSStefano Zampini } 107808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 107908122e43SStefano Zampini if (B_ierr) { 108008122e43SStefano Zampini if (B_ierr < 0 ) { 108108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 108208122e43SStefano Zampini } else if (B_ierr <= B_N) { 108308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 108408122e43SStefano Zampini } else { 10859552c7c7SStefano 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); 108608122e43SStefano Zampini } 108708122e43SStefano Zampini } 108808122e43SStefano Zampini 108908122e43SStefano Zampini if (B_neigs > nmax) { 1090fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1091fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1092fd14bc51SStefano Zampini } 1093f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 109408122e43SStefano Zampini B_neigs = nmax; 109508122e43SStefano Zampini } 109608122e43SStefano Zampini 10979552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 10989552c7c7SStefano Zampini if (B_neigs < nmin_s) { 109908122e43SStefano Zampini PetscBLASInt B_neigs2; 110008122e43SStefano Zampini 1101f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1102f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1103fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1104fd14bc51SStefano 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); 1105fd14bc51SStefano Zampini } 11069ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 110708122e43SStefano Zampini PetscInt j; 110808122e43SStefano Zampini for (j=0;j<subset_size;j++) { 110908122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 111008122e43SStefano Zampini } 111108122e43SStefano Zampini for (j=0;j<subset_size;j++) { 111208122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 111308122e43SStefano Zampini } 111408122e43SStefano Zampini } else { 111508122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111608122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111708122e43SStefano Zampini } 111808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 111908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1120f6f667cfSStefano 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)); 112108122e43SStefano Zampini #else 1122f6f667cfSStefano 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)); 112308122e43SStefano Zampini #endif 112408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 112508122e43SStefano Zampini B_neigs += B_neigs2; 112608122e43SStefano Zampini } 112708122e43SStefano Zampini if (B_ierr) { 112808122e43SStefano Zampini if (B_ierr < 0 ) { 112908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 113008122e43SStefano Zampini } else if (B_ierr <= B_N) { 113108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 113208122e43SStefano Zampini } else { 11339552c7c7SStefano 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); 113408122e43SStefano Zampini } 113508122e43SStefano Zampini } 1136fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1137ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 113808122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 113908122e43SStefano Zampini if (eigs[j] == 0.0) { 1140ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 114108122e43SStefano Zampini } else { 1142ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1143fd14bc51SStefano Zampini } 114408122e43SStefano Zampini } 114508122e43SStefano Zampini } 114608122e43SStefano Zampini } else { 114708122e43SStefano Zampini /* TODO */ 114808122e43SStefano Zampini } 1149aff50787SStefano Zampini } 11508bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 11518bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 11529162d606SStefano Zampini if (B_neigs) { 11539162d606SStefano 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); 1154fd14bc51SStefano Zampini 1155fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 11569552c7c7SStefano Zampini PetscInt ii; 11579552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1158ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11599552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1160ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1161ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1162ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1163ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1164ac47001eSStefano Zampini #else 1165ac47001eSStefano 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); 1166ac47001eSStefano Zampini #endif 11679552c7c7SStefano Zampini } 11689552c7c7SStefano Zampini } 1169fd14bc51SStefano Zampini } 117008122e43SStefano Zampini #if 0 11719162d606SStefano Zampini for (j=0;j<B_neigs;j++) { 117208122e43SStefano Zampini PetscBLASInt Blas_N,Blas_one = 1.0; 117308122e43SStefano Zampini PetscScalar norm; 117408122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&Blas_N);CHKERRQ(ierr); 11759162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size, 11769162d606SStefano Zampini &Blas_one,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 117708122e43SStefano Zampini if (pcbddc->adaptive_constraints_data[cum2] > 0.0) { 117808122e43SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 117908122e43SStefano Zampini } else { 118008122e43SStefano Zampini norm = -1.0/PetscSqrtReal(PetscRealPart(norm)); 118108122e43SStefano Zampini } 11829162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum]+j*subset_size,&Blas_one)); 1183b1b3d7a2SStefano Zampini } 1184b1b3d7a2SStefano Zampini #endif 11859162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 11869162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 11879162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 11889162d606SStefano Zampini cum++; 118908122e43SStefano Zampini } 119008122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 119108122e43SStefano Zampini /* shift for next computation */ 119208122e43SStefano Zampini cumarray += subset_size*subset_size; 119308122e43SStefano Zampini } 1194fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1195fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1196fd14bc51SStefano Zampini } 119708122e43SStefano Zampini 119808122e43SStefano Zampini if (mss) { 119908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 120008122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1201f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1202f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1203f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 120408122e43SStefano Zampini } 1205f6f667cfSStefano Zampini if (allocated_S_St) { 1206f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1207f6f667cfSStefano Zampini } 1208f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 120908122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 121008122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 121108122e43SStefano Zampini #endif 121208122e43SStefano Zampini if (pcbddc->dbg_flag) { 12131b968477SStefano Zampini PetscInt maxneigs_r; 121408122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12159b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 121608122e43SStefano Zampini } 121708122e43SStefano Zampini PetscFunctionReturn(0); 121808122e43SStefano Zampini } 1219b1b3d7a2SStefano Zampini 1220674ae819SStefano Zampini #undef __FUNCT__ 1221c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1222c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1223c8587f34SStefano Zampini { 1224c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12258629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1226c8587f34SStefano Zampini PetscErrorCode ierr; 1227c8587f34SStefano Zampini 1228c8587f34SStefano Zampini PetscFunctionBegin; 1229f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12305e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1231c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1232c8587f34SStefano Zampini 1233684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12340fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1235684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1236c8587f34SStefano Zampini 1237c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1238b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1239c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1240c8587f34SStefano Zampini } 1241c8587f34SStefano Zampini 12428629588bSStefano Zampini /* 12438629588bSStefano Zampini Setup local correction and local part of coarse basis. 12448629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12458629588bSStefano Zampini */ 124647f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12478629588bSStefano Zampini 12488629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12498629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12508629588bSStefano Zampini 12518629588bSStefano Zampini /* free */ 12528629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1253c8587f34SStefano Zampini PetscFunctionReturn(0); 1254c8587f34SStefano Zampini } 1255c8587f34SStefano Zampini 1256c8587f34SStefano Zampini #undef __FUNCT__ 1257674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1258674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1259674ae819SStefano Zampini { 1260674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1261674ae819SStefano Zampini PetscErrorCode ierr; 1262674ae819SStefano Zampini 1263674ae819SStefano Zampini PetscFunctionBegin; 1264674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1265674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 126630368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1267674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1268674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1269785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1270674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1271f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1272f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1273785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 127463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 127563602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1276674ae819SStefano Zampini PetscFunctionReturn(0); 1277674ae819SStefano Zampini } 1278674ae819SStefano Zampini 1279674ae819SStefano Zampini #undef __FUNCT__ 1280674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1281674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1282674ae819SStefano Zampini { 1283674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12844f1b2e48SStefano Zampini PetscInt i; 1285674ae819SStefano Zampini PetscErrorCode ierr; 1286674ae819SStefano Zampini 1287674ae819SStefano Zampini PetscFunctionBegin; 1288b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1289674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1290674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1291674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 12924f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 12934f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 12944f1b2e48SStefano Zampini } 12954f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1296b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1297674ae819SStefano Zampini PetscFunctionReturn(0); 1298674ae819SStefano Zampini } 1299674ae819SStefano Zampini 1300674ae819SStefano Zampini #undef __FUNCT__ 1301674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1302674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1303674ae819SStefano Zampini { 1304674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1305674ae819SStefano Zampini PetscErrorCode ierr; 1306674ae819SStefano Zampini 1307674ae819SStefano Zampini PetscFunctionBegin; 1308674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 130958da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1310ca92afb2SStefano Zampini PetscScalar *array; 131106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 131206656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 131358da7f69SStefano Zampini } 1314674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1315674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 131615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 131715aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1318674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1319674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1320674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 132106656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1322674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1323674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13248ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1325674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1326674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1327674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1328f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1329f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1330f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1331f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1332727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13330e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1334f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 133570cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 13366e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 133781d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 13380369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 13398b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 13404f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13418b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1342ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1343ca92afb2SStefano Zampini PetscInt i; 1344ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1345ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1346ca92afb2SStefano Zampini } 1347ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1348ca92afb2SStefano Zampini } 13494f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1350674ae819SStefano Zampini PetscFunctionReturn(0); 1351674ae819SStefano Zampini } 1352674ae819SStefano Zampini 1353674ae819SStefano Zampini #undef __FUNCT__ 1354f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1355f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 13566bfb1811SStefano Zampini { 13576bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13586bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 13596bfb1811SStefano Zampini VecType impVecType; 13604f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 13616bfb1811SStefano Zampini PetscErrorCode ierr; 13626bfb1811SStefano Zampini 13636bfb1811SStefano Zampini PetscFunctionBegin; 1364f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1365019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1366f4ddd8eeSStefano Zampini } 1367e7b262bdSStefano Zampini /* get sizes */ 13684f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1369b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 13706bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1371e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1372e7b262bdSStefano Zampini /* R nodes */ 1373e7b262bdSStefano Zampini old_size = -1; 1374e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1375e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1376e7b262bdSStefano Zampini } 1377e7b262bdSStefano Zampini if (n_R != old_size) { 1378e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1379e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13806bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 13816bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 13826bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 13836bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1384e7b262bdSStefano Zampini } 1385e7b262bdSStefano Zampini /* local primal dofs */ 1386e7b262bdSStefano Zampini old_size = -1; 1387e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1388e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1389e7b262bdSStefano Zampini } 1390e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1391e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 139283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1393e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 13946bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1395e7b262bdSStefano Zampini } 1396e7b262bdSStefano Zampini /* local explicit constraints */ 1397e7b262bdSStefano Zampini old_size = -1; 1398e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1399e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1400e7b262bdSStefano Zampini } 1401e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1402e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 140383b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 140483b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 140583b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 140683b7ccabSStefano Zampini } 14076bfb1811SStefano Zampini PetscFunctionReturn(0); 14086bfb1811SStefano Zampini } 14096bfb1811SStefano Zampini 14106bfb1811SStefano Zampini #undef __FUNCT__ 141147f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 141247f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 141388ebb749SStefano Zampini { 141425084f0cSStefano Zampini PetscErrorCode ierr; 141525084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 141688ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 141788ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1418d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 141925084f0cSStefano Zampini /* submatrices of local problem */ 142080677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 1421*fc227af8SStefano Zampini /* submatrices for benign trick */ 1422*fc227af8SStefano Zampini Mat B0_R = NULL, B0_V = NULL; 142306656605SStefano Zampini /* submatrices of local coarse problem */ 142406656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 142525084f0cSStefano Zampini /* working matrices */ 142606656605SStefano Zampini Mat C_CR; 142725084f0cSStefano Zampini /* additional working stuff */ 142806656605SStefano Zampini PC pc_R; 14294f1b2e48SStefano Zampini Mat F; 1430a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 143106656605SStefano Zampini 143225084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 143306656605SStefano Zampini PetscScalar *work; 143406656605SStefano Zampini PetscInt *idx_V_B; 1435ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 143606656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1437ffd830a3SStefano Zampini 143825084f0cSStefano Zampini /* some shortcuts to scalars */ 143906656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 144088ebb749SStefano Zampini 144188ebb749SStefano Zampini PetscFunctionBegin; 1442ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1443ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1444ffd830a3SStefano Zampini } 1445ffd830a3SStefano Zampini 1446ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1447b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14484f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1449b371cd4fSStefano Zampini n_B = pcis->n_B; 1450b371cd4fSStefano Zampini n_D = pcis->n - n_B; 145188ebb749SStefano Zampini n_R = pcis->n - n_vertices; 145288ebb749SStefano Zampini 145388ebb749SStefano Zampini /* vertices in boundary numbering */ 1454785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 14550e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 145688ebb749SStefano Zampini if (i != n_vertices) { 145722d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 145888ebb749SStefano Zampini } 145988ebb749SStefano Zampini 146006656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1461019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 146206656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 146306656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 146406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 146506656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 146606656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 146706656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 146806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 146906656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 147006656605SStefano Zampini 147106656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 147206656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 147306656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 147406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 147506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1476ffd830a3SStefano Zampini lda_rhs = n_R; 1477a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 147806656605SStefano Zampini if (isLU || isILU || isCHOL) { 147906656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1480df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1481df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1482d62866d3SStefano Zampini MatFactorType type; 1483d62866d3SStefano Zampini 1484df4d28bfSStefano Zampini F = reuse_solver->F; 14856816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1486d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1487ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1488df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 148906656605SStefano Zampini } else { 149006656605SStefano Zampini F = NULL; 149106656605SStefano Zampini } 149206656605SStefano Zampini 1493ffd830a3SStefano Zampini /* allocate workspace */ 1494ffd830a3SStefano Zampini n = 0; 1495ffd830a3SStefano Zampini if (n_constraints) { 1496ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1497ffd830a3SStefano Zampini } 1498ffd830a3SStefano Zampini if (n_vertices) { 1499ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1500ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1501ffd830a3SStefano Zampini } 1502ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1503ffd830a3SStefano Zampini 150488ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 150588ebb749SStefano Zampini if (n_constraints) { 150606656605SStefano Zampini Mat M1,M2,M3; 150780677318SStefano Zampini Mat auxmat; 150806656605SStefano Zampini IS is_aux; 150980677318SStefano Zampini PetscScalar *array,*array2; 151006656605SStefano Zampini 1511f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 151280677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 151388ebb749SStefano Zampini 151425084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 151525084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15168ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 151780677318SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&auxmat);CHKERRQ(ierr); 151888ebb749SStefano Zampini 151980677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 152080677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1521ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 152288ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 152306656605SStefano Zampini const PetscScalar *row_cmat_values; 152406656605SStefano Zampini const PetscInt *row_cmat_indices; 152506656605SStefano Zampini PetscInt size_of_constraint,j; 152688ebb749SStefano Zampini 152706656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 152806656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1529ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 153006656605SStefano Zampini } 153106656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 153206656605SStefano Zampini } 1533ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 153406656605SStefano Zampini if (F) { 153506656605SStefano Zampini Mat B; 153606656605SStefano Zampini 1537ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1538a3df083aSStefano Zampini if (need_benign_correction) { 1539a3df083aSStefano Zampini PetscScalar *marr; 1540df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1541a3df083aSStefano Zampini 1542a3df083aSStefano Zampini ierr = MatDenseGetArray(B,&marr);CHKERRQ(ierr); 1543a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1544a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1545df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1546a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1547a3df083aSStefano Zampini } 1548a3df083aSStefano Zampini ierr = MatDenseRestoreArray(B,&marr);CHKERRQ(ierr); 1549a3df083aSStefano Zampini } 155080677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1551a3df083aSStefano Zampini if (need_benign_correction) { 1552a3df083aSStefano Zampini PetscScalar *marr; 1553df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1554a3df083aSStefano Zampini 1555a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1556a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1557a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1558df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1559a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1560a3df083aSStefano Zampini } 1561a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1562a3df083aSStefano Zampini } 156306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 156406656605SStefano Zampini } else { 156580677318SStefano Zampini PetscScalar *marr; 156680677318SStefano Zampini 156780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 156806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1569ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1570ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 157106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 157206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 157306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 157406656605SStefano Zampini } 157580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 157606656605SStefano Zampini } 157780677318SStefano Zampini if (!pcbddc->switch_static) { 157880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 157980677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 158080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 158180677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1582ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 158380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 158480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 158580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 158680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 158780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 158880677318SStefano Zampini } 158980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 159080677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 159180677318SStefano Zampini ierr = MatMatMult(auxmat,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 159280677318SStefano Zampini } else { 1593ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1594ffd830a3SStefano Zampini IS dummy; 1595ffd830a3SStefano Zampini 1596ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 1597ffd830a3SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,dummy,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1598ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1599ffd830a3SStefano Zampini } else { 160080677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 160180677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1602ffd830a3SStefano Zampini } 160325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 160480677318SStefano Zampini } 160580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 160680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 160780677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 160806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 160906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 161080677318SStefano Zampini if (isCHOL) { 161180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 161280677318SStefano Zampini } else { 161325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 161480677318SStefano Zampini } 161580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 161606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 161725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 161825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 161925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 162080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 162180677318SStefano Zampini ierr = MatMatMult(M1,auxmat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 162280677318SStefano Zampini ierr = MatDestroy(&auxmat);CHKERRQ(ierr); 162306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 162406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1625f4ddd8eeSStefano Zampini } 1626*fc227af8SStefano Zampini 1627*fc227af8SStefano Zampini /* Extract B0_R */ 16284f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1629d16cbb6bSStefano Zampini IS dummy; 1630d16cbb6bSStefano Zampini PetscReal norm; 1631d16cbb6bSStefano Zampini 16324f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16334f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&B0_R);CHKERRQ(ierr); 1634d16cbb6bSStefano Zampini ierr = MatNorm(B0_R,NORM_INFINITY,&norm);CHKERRQ(ierr); 1635*fc227af8SStefano Zampini if (norm < PETSC_SMALL) { 1636d16cbb6bSStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 1637*fc227af8SStefano Zampini } 1638d16cbb6bSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1639d16cbb6bSStefano Zampini } 1640d16cbb6bSStefano Zampini 1641*fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 164288ebb749SStefano Zampini if (n_vertices) { 164306656605SStefano Zampini IS is_aux; 16443a50541eSStefano Zampini 1645df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16466816873aSStefano Zampini IS tis; 16476816873aSStefano Zampini 16486816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16496816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16506816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16516816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16526816873aSStefano Zampini } else { 16533a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16546816873aSStefano Zampini } 16559577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16569577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 165704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 16584f1b2e48SStefano Zampini if (pcbddc->benign_n) { 1659019a44ceSStefano Zampini IS dummy; 1660019a44ceSStefano Zampini 16614f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 16624f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,is_aux,MAT_INITIAL_MATRIX,&B0_V);CHKERRQ(ierr); 1663019a44ceSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1664019a44ceSStefano Zampini } 166525084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 166688ebb749SStefano Zampini } 166788ebb749SStefano Zampini 166888ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1669f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 167006656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 167106656605SStefano Zampini if (pcbddc->coarse_phi_D) { 167206656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 167306656605SStefano Zampini } 1674f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 167506656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 167606656605SStefano Zampini PetscScalar *marray; 167706656605SStefano Zampini 167806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 167906656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1680f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1681f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1682f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1683f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1684f4ddd8eeSStefano Zampini } 1685f4ddd8eeSStefano Zampini } 168606656605SStefano Zampini 1687f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 168806656605SStefano Zampini PetscScalar *marray; 168988ebb749SStefano Zampini 169006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16918eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 169206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 169388ebb749SStefano Zampini } 16943301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 169506656605SStefano Zampini n *= 2; 169688ebb749SStefano Zampini } 169706656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 169806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 169906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17008eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 170106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 170206656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 170388ebb749SStefano Zampini } 17043301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 170506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17068eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 170706656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 170806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 170988ebb749SStefano Zampini } 171088ebb749SStefano Zampini } else { 1711c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1712c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17131b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1714c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1715c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1716c0553b1fSStefano Zampini } 171788ebb749SStefano Zampini } 171806656605SStefano Zampini } 1719019a44ceSStefano Zampini 172006656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17214f1b2e48SStefano Zampini p0_lidx_I = NULL; 17224f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1723d12edf2fSStefano Zampini const PetscInt *idxs; 1724d12edf2fSStefano Zampini 1725d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17264f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17274f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17284f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17294f1b2e48SStefano Zampini } 1730d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1731d12edf2fSStefano Zampini } 1732d16cbb6bSStefano Zampini 173306656605SStefano Zampini /* vertices */ 173406656605SStefano Zampini if (n_vertices) { 173516f15bc4SStefano Zampini 1736ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 173704708bb6SStefano Zampini 173816f15bc4SStefano Zampini if (n_R) { 173914393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 174006656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 174116f15bc4SStefano Zampini PetscScalar *x,*y; 174204708bb6SStefano Zampini PetscBool isseqaij; 174306656605SStefano Zampini 174421eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 174514393ed6SStefano Zampini if (need_benign_correction) { 174614393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 174714393ed6SStefano Zampini IS is_p0; 174814393ed6SStefano Zampini PetscInt *idxs_p0,n; 174914393ed6SStefano Zampini 175014393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 175114393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 175214393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 175314393ed6SStefano Zampini if (n != pcbddc->benign_n) { 175414393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 175514393ed6SStefano Zampini } 175614393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 175714393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 175814393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 175914393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 176014393ed6SStefano Zampini } 176114393ed6SStefano Zampini 1762ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1763ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1764ffd830a3SStefano Zampini } else { 1765ca92afb2SStefano Zampini PetscScalar *av,*array; 1766ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1767ca92afb2SStefano Zampini PetscInt n; 1768ca92afb2SStefano Zampini PetscBool flg_row; 1769ffd830a3SStefano Zampini 1770ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1771ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1772ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1773ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1774ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1775ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1776ca92afb2SStefano Zampini PetscInt j; 1777ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1778ffd830a3SStefano Zampini } 1779ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1780ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1781ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1782ffd830a3SStefano Zampini } 1783ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 178414393ed6SStefano Zampini if (F) { 1785a3df083aSStefano Zampini if (need_benign_correction) { 1786df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1787a3df083aSStefano Zampini PetscScalar *marr; 1788a3df083aSStefano Zampini 1789a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 179014393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 179114393ed6SStefano Zampini 179214393ed6SStefano Zampini | 0 0 0 | (V) 179314393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 179414393ed6SStefano Zampini | 0 0 -1 | (p0) 179514393ed6SStefano Zampini 179614393ed6SStefano Zampini */ 1797df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 179814393ed6SStefano Zampini const PetscScalar *vals; 179914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 180014393ed6SStefano Zampini PetscInt n,j,nz; 180114393ed6SStefano Zampini 1802df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1803df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 180414393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 180514393ed6SStefano Zampini for (j=0;j<n;j++) { 180614393ed6SStefano Zampini PetscScalar val = vals[j]; 180714393ed6SStefano Zampini PetscInt k,col = idxs[j]; 180814393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 180914393ed6SStefano Zampini } 181014393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1811df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 181214393ed6SStefano Zampini } 181314393ed6SStefano Zampini /* need to correct the rhs */ 1814a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1815a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1816df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1817a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1818a3df083aSStefano Zampini } 1819a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1820a3df083aSStefano Zampini } 182106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 182214393ed6SStefano Zampini /* need to correct the solution */ 1823a3df083aSStefano Zampini if (need_benign_correction) { 1824df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1825a3df083aSStefano Zampini PetscScalar *marr; 1826a3df083aSStefano Zampini 1827a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1828a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1829a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1830df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1831a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1832a3df083aSStefano Zampini } 1833a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1834a3df083aSStefano Zampini } 183506656605SStefano Zampini } else { 183606656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 183706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1838ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1839ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 184006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 184106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 184206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 184306656605SStefano Zampini } 184406656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 184506656605SStefano Zampini } 184680677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1847ffd830a3SStefano Zampini /* S_VV and S_CV */ 184806656605SStefano Zampini if (n_constraints) { 184906656605SStefano Zampini Mat B; 185080677318SStefano Zampini 1851ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 185280677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1853ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1854ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 185580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 185680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 185780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 185880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 185980677318SStefano Zampini } 1860ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 186180677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 186280677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1863ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 186480677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 186506656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1866ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1867ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 186806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 186906656605SStefano Zampini } 187004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 187104708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 187204708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 187304708bb6SStefano Zampini } 1874ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1875ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1876ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1877ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1878ffd830a3SStefano Zampini } 187906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 188014393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 188114393ed6SStefano Zampini if (need_benign_correction) { 1882df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 188314393ed6SStefano Zampini PetscScalar *marr,*sums; 188414393ed6SStefano Zampini 188514393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 188614393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1887df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 188814393ed6SStefano Zampini const PetscScalar *vals; 188914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 189014393ed6SStefano Zampini PetscInt n,j,nz; 189114393ed6SStefano Zampini 1892df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1893df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 189414393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 189514393ed6SStefano Zampini PetscInt k; 189614393ed6SStefano Zampini sums[j] = 0.; 189714393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 189814393ed6SStefano Zampini } 189914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 190014393ed6SStefano Zampini for (j=0;j<n;j++) { 190114393ed6SStefano Zampini PetscScalar val = vals[j]; 190214393ed6SStefano Zampini PetscInt k; 190314393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 190414393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 190514393ed6SStefano Zampini } 190614393ed6SStefano Zampini } 190714393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1908df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 190914393ed6SStefano Zampini } 191014393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 191114393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 191214393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 191314393ed6SStefano Zampini } 191480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 191506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 191606656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 191706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 191806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 191906656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 192006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 192106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1922d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1923019a44ceSStefano Zampini } else { 1924d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1925d16cbb6bSStefano Zampini } 19264f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1927019a44ceSStefano Zampini const PetscScalar *vals; 1928019a44ceSStefano Zampini const PetscInt *idxs; 19294f1b2e48SStefano Zampini PetscInt n,j,primal_idx; 1930019a44ceSStefano Zampini 19314f1b2e48SStefano Zampini ierr = MatGetRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 19324f1b2e48SStefano Zampini primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + i; 1933d16cbb6bSStefano Zampini for (j=0;j<n;j++) { 19344f1b2e48SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+idxs[j]] = vals[j]; 19354f1b2e48SStefano Zampini coarse_submat_vals[idxs[j]*pcbddc->local_primal_size+primal_idx] = vals[j]; 1936019a44ceSStefano Zampini } 19374f1b2e48SStefano Zampini ierr = MatRestoreRow(B0_V,i,&n,&idxs,&vals);CHKERRQ(ierr); 193816f15bc4SStefano Zampini } 193921eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1940d16cbb6bSStefano Zampini 194106656605SStefano Zampini /* coarse basis functions */ 194206656605SStefano Zampini for (i=0;i<n_vertices;i++) { 194316f15bc4SStefano Zampini PetscScalar *y; 194416f15bc4SStefano Zampini 1945ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 194606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 194706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 194806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 194906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 195006656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 195106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 195206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 195306656605SStefano Zampini 195406656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19554f1b2e48SStefano Zampini PetscInt j; 19564f1b2e48SStefano Zampini 195706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 195806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 195906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 196106656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19624f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 196306656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 196406656605SStefano Zampini } 196506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 196606656605SStefano Zampini } 196704708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 196804708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 196906656605SStefano Zampini } 197006656605SStefano Zampini 197106656605SStefano Zampini if (n_constraints) { 197206656605SStefano Zampini Mat B; 197306656605SStefano Zampini 1974ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 197506656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 197680677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 197706656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 197806656605SStefano Zampini if (n_vertices) { 197980677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 198080677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 198180677318SStefano Zampini } else { 198280677318SStefano Zampini Mat S_VCt; 198380677318SStefano Zampini 1984ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1985ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1986ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&B);CHKERRQ(ierr); 1987ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1988ffd830a3SStefano Zampini } 198980677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 199080677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 199180677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 199280677318SStefano Zampini } 199306656605SStefano Zampini } 199406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 199506656605SStefano Zampini /* coarse basis functions */ 199606656605SStefano Zampini for (i=0;i<n_constraints;i++) { 199706656605SStefano Zampini PetscScalar *y; 199806656605SStefano Zampini 1999ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 200006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 200206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 200506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 200606656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20074f1b2e48SStefano Zampini PetscInt j; 20084f1b2e48SStefano Zampini 200906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 201006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 201106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 201306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20144f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 201506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 201606656605SStefano Zampini } 2017*fc227af8SStefano Zampini if (B0_R) { 2018*fc227af8SStefano Zampini Vec benign_proj; 2019*fc227af8SStefano Zampini const PetscScalar *array; 2020*fc227af8SStefano Zampini PetscInt j; 2021*fc227af8SStefano Zampini 2022*fc227af8SStefano Zampini ierr = MatCreateVecs(B0_R,NULL,&benign_proj);CHKERRQ(ierr); 2023*fc227af8SStefano Zampini ierr = MatMult(B0_R,pcbddc->vec1_R,benign_proj);CHKERRQ(ierr); 2024*fc227af8SStefano Zampini ierr = VecGetArrayRead(benign_proj,&array);CHKERRQ(ierr); 2025*fc227af8SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 2026*fc227af8SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2027*fc227af8SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i+n_vertices] = array[j]; 2028*fc227af8SStefano Zampini coarse_submat_vals[(i+n_vertices)*pcbddc->local_primal_size+primal_idx] = array[j]; 2029*fc227af8SStefano Zampini } 2030*fc227af8SStefano Zampini ierr = VecRestoreArrayRead(benign_proj,&array);CHKERRQ(ierr); 2031*fc227af8SStefano Zampini ierr = VecDestroy(&benign_proj);CHKERRQ(ierr); 2032*fc227af8SStefano Zampini } 203306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 203406656605SStefano Zampini } 203506656605SStefano Zampini } 203680677318SStefano Zampini if (n_constraints) { 203780677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 203880677318SStefano Zampini } 20394f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 2040019a44ceSStefano Zampini ierr = MatDestroy(&B0_V);CHKERRQ(ierr); 2041*fc227af8SStefano Zampini ierr = MatDestroy(&B0_R);CHKERRQ(ierr); 2042019a44ceSStefano Zampini 204306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 20443301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2045ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2046ffd830a3SStefano Zampini PetscScalar *marray; 204706656605SStefano Zampini 204806656605SStefano Zampini if (n_constraints) { 2049ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 205006656605SStefano Zampini 2051ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 205206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2053ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 205416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 205506656605SStefano Zampini if (n_vertices) { 2056ffd830a3SStefano Zampini Mat S_VCT; 205706656605SStefano Zampini 205806656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2059ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 206016f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 206106656605SStefano Zampini } 2062ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 206306656605SStefano Zampini } 206416f15bc4SStefano Zampini if (n_vertices && n_R) { 2065ffd830a3SStefano Zampini PetscScalar *av,*marray; 2066ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2067ffd830a3SStefano Zampini PetscInt n; 2068ffd830a3SStefano Zampini PetscBool flg_row; 206906656605SStefano Zampini 2070ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2071ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2072ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2073ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2074ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2075ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2076ffd830a3SStefano Zampini PetscInt j; 2077ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2078ffd830a3SStefano Zampini } 2079ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2080ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2081ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 208206656605SStefano Zampini } 208306656605SStefano Zampini 2084ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2085ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2086ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2087ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2088ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 208906656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 209006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 209106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 209206656605SStefano Zampini } 2093ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2094ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2095ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2096ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2097ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2098ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2099ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2100ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 210106656605SStefano Zampini } 2102ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 210306656605SStefano Zampini /* coarse basis functions */ 210406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 210506656605SStefano Zampini PetscScalar *y; 210606656605SStefano Zampini 2107ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 210806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 210906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 211006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 211106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 211206656605SStefano Zampini if (i<n_vertices) { 211306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 211406656605SStefano Zampini } 211506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 211606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 211706656605SStefano Zampini 211806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 211906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 212006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 212106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 212206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 212306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 212406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 212506656605SStefano Zampini } 212606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 212706656605SStefano Zampini } 2128ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2129ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 213006656605SStefano Zampini } 2131d62866d3SStefano Zampini /* free memory */ 213288ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 213306656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 213406656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 213506656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 213606656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2137d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2138d62866d3SStefano Zampini if (n_vertices) { 2139d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2140d62866d3SStefano Zampini } 2141d62866d3SStefano Zampini if (n_constraints) { 2142d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2143d62866d3SStefano Zampini } 214488ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 214588ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 214688ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2147d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 214888ebb749SStefano Zampini Mat coarse_sub_mat; 214925084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 215088ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 215188ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 215288ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 21538bec7fa6SStefano Zampini Mat C_B,CPHI; 21548bec7fa6SStefano Zampini IS is_dummy; 21558bec7fa6SStefano Zampini Vec mones; 215688ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 215788ebb749SStefano Zampini PetscReal real_value; 215888ebb749SStefano Zampini 2159a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2160a3df083aSStefano Zampini Mat A; 2161a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2162a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2163a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2164a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2165a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2166a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2167a3df083aSStefano Zampini } else { 216888ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 216988ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 217088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 217188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2172a3df083aSStefano Zampini } 217388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 217488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2175ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 217688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 217788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 217888ebb749SStefano Zampini } 217988ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 218088ebb749SStefano Zampini 218125084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 21823301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 218325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2184ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 218588ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 218688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 218788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 218888ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 218988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 219088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 219188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 219288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 219388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 219488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 219588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 219688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 219788ebb749SStefano Zampini } else { 219888ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 219988ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 220088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 220188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 220288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 220388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 220488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 220588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 220688ebb749SStefano Zampini } 220788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 220888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 220988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 221088ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 22114f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2212*fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2213d12edf2fSStefano Zampini PetscScalar *data,*data2; 22144f1b2e48SStefano Zampini PetscInt j; 2215d12edf2fSStefano Zampini 22164f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2217*fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2218d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2219d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2220d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2221d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 22224f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 22234f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2224d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 22254f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 22264f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22274f1b2e48SStefano Zampini } 2228d12edf2fSStefano Zampini } 2229d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2230d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2231d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2232d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2233d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2234d12edf2fSStefano Zampini } 2235d12edf2fSStefano Zampini #if 0 2236d12edf2fSStefano Zampini { 2237d12edf2fSStefano Zampini PetscViewer viewer; 2238d12edf2fSStefano Zampini char filename[256]; 2239ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2240d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2241d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2242ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2243ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2244ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2245d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 2246ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2247ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 2248ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2249ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2250ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2251ffd830a3SStefano Zampini } 2252ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2253ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2254ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2255ffd830a3SStefano Zampini } 2256ffd830a3SStefano Zampini if (pcbddc->coarse_phi_B) { 2257ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2258ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2259ffd830a3SStefano Zampini } 2260d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2261d12edf2fSStefano Zampini } 2262d12edf2fSStefano Zampini #endif 226381d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 22648bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 22651575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 226606656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 22678bec7fa6SStefano Zampini 22688bec7fa6SStefano Zampini /* check constraints */ 22694f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 22708bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 22718bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 22728bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 22738bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 22748bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 22758bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 22768bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2277bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2278ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2279bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2280bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2281bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2282bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2283bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 228488ebb749SStefano Zampini } 22858bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 22868bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 22878bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 22888bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2289d12edf2fSStefano Zampini } 229025084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 229188ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 229288ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 229388ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 229488ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 229588ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 229688ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 229788ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 229888ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 229988ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 230088ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2301ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 230288ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 230388ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 230488ebb749SStefano Zampini } 230588ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 230688ebb749SStefano Zampini } 23078629588bSStefano Zampini /* get back data */ 23088629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 230988ebb749SStefano Zampini PetscFunctionReturn(0); 231088ebb749SStefano Zampini } 231188ebb749SStefano Zampini 231288ebb749SStefano Zampini #undef __FUNCT__ 2313d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2314d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2315aa0d41d4SStefano Zampini { 2316d65f70fdSStefano Zampini Mat *work_mat; 2317d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2318d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2319d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2320aa0d41d4SStefano Zampini PetscErrorCode ierr; 2321aa0d41d4SStefano Zampini 2322aa0d41d4SStefano Zampini PetscFunctionBegin; 2323d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2324d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2325d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2326d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2327aa0d41d4SStefano Zampini 2328d65f70fdSStefano Zampini if (!rsorted) { 2329906d46d4SStefano Zampini const PetscInt *idxs; 2330906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2331aa0d41d4SStefano Zampini 2332d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2333d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2334d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2335d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2336aa0d41d4SStefano Zampini } 2337d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2338d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2339d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2340d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2341aa0d41d4SStefano Zampini } 2342d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2343d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2344d65f70fdSStefano Zampini } else { 2345d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2346d65f70fdSStefano Zampini isrow_s = isrow; 2347aa0d41d4SStefano Zampini } 2348906d46d4SStefano Zampini 2349d65f70fdSStefano Zampini if (!csorted) { 2350d65f70fdSStefano Zampini if (isrow == iscol) { 2351d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2352d65f70fdSStefano Zampini iscol_s = isrow_s; 2353d65f70fdSStefano Zampini } else { 2354d65f70fdSStefano Zampini const PetscInt *idxs; 2355d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2356906d46d4SStefano Zampini 2357d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2358d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2359d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2360d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2361d65f70fdSStefano Zampini } 2362d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2363d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2364d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2365d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2366d65f70fdSStefano Zampini } 2367d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2368d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2369d65f70fdSStefano Zampini } 2370d65f70fdSStefano Zampini } else { 2371d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2372d65f70fdSStefano Zampini iscol_s = iscol; 2373d65f70fdSStefano Zampini } 2374d65f70fdSStefano Zampini 2375d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2376d65f70fdSStefano Zampini 2377d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2378906d46d4SStefano Zampini Mat new_mat; 2379d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2380906d46d4SStefano Zampini 2381d65f70fdSStefano Zampini if (!rsorted) { 2382d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2383d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2384d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2385d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2386906d46d4SStefano Zampini } 2387d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2388d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2389d65f70fdSStefano Zampini } else { 2390d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2391906d46d4SStefano Zampini } 2392d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2393d65f70fdSStefano Zampini 2394d65f70fdSStefano Zampini if (!csorted) { 2395d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2396d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2397d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2398d65f70fdSStefano Zampini } else { 2399d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2400d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2401d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2402d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2403d65f70fdSStefano Zampini } 2404d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2405d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2406d65f70fdSStefano Zampini } 2407d65f70fdSStefano Zampini } else { 2408d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2409d65f70fdSStefano Zampini } 2410d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2411d65f70fdSStefano Zampini 2412d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2413d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2414d65f70fdSStefano Zampini work_mat[0] = new_mat; 2415d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2416d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2417d65f70fdSStefano Zampini } 2418d65f70fdSStefano Zampini 2419d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2420d65f70fdSStefano Zampini *B = work_mat[0]; 2421d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2422d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2423d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2424d65f70fdSStefano Zampini PetscFunctionReturn(0); 2425d65f70fdSStefano Zampini } 2426d65f70fdSStefano Zampini 2427d65f70fdSStefano Zampini #undef __FUNCT__ 24285e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 24295e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2430aa0d41d4SStefano Zampini { 2431aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 24325e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2433d65f70fdSStefano Zampini Mat new_mat; 24345e8657edSStefano Zampini IS is_local,is_global; 2435d65f70fdSStefano Zampini PetscInt local_size; 2436d65f70fdSStefano Zampini PetscBool isseqaij; 2437aa0d41d4SStefano Zampini PetscErrorCode ierr; 2438aa0d41d4SStefano Zampini 2439aa0d41d4SStefano Zampini PetscFunctionBegin; 2440aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 24415e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 24425e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2443b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2444aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2445d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2446aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2447906d46d4SStefano Zampini 2448906d46d4SStefano Zampini /* check */ 2449906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2450906d46d4SStefano Zampini Vec x,x_change; 2451906d46d4SStefano Zampini PetscReal error; 2452906d46d4SStefano Zampini 24535e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2454906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 24555e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2456e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2457e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2458d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2459e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2460e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2461906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2462906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2463906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2464906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2465906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2466906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2467906d46d4SStefano Zampini } 2468906d46d4SStefano Zampini 246922d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 24709b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 247122d5777bSStefano Zampini if (isseqaij) { 24721cf9b237SStefano Zampini Mat M; 24731cf9b237SStefano Zampini 24741cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24751cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24761cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2477aa0d41d4SStefano Zampini } else { 24781cf9b237SStefano Zampini Mat work_mat,M; 24791cf9b237SStefano Zampini 2480aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 24811cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24821cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24831cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2484aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2485aa0d41d4SStefano Zampini } 24863301b35fSStefano Zampini if (matis->A->symmetric_set) { 24873301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2488e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24893301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2490e496cd5dSStefano Zampini #endif 24913301b35fSStefano Zampini } 249245a1bb75SStefano Zampini /* 249345a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2494d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 249545a1bb75SStefano Zampini */ 2496d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2497aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2498aa0d41d4SStefano Zampini } 2499aa0d41d4SStefano Zampini 2500aa0d41d4SStefano Zampini #undef __FUNCT__ 2501a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25028ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2503a64d13efSStefano Zampini { 2504a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2505a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2506d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 250753892102SStefano Zampini PetscInt *idx_R_local=NULL; 25083a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25093a50541eSStefano Zampini PetscInt vbs,bs; 25106816873aSStefano Zampini PetscBT bitmask=NULL; 2511a64d13efSStefano Zampini PetscErrorCode ierr; 2512a64d13efSStefano Zampini 2513a64d13efSStefano Zampini PetscFunctionBegin; 2514b23d619eSStefano Zampini /* 2515b23d619eSStefano Zampini No need to setup local scatters if 2516b23d619eSStefano Zampini - primal space is unchanged 2517b23d619eSStefano Zampini AND 2518b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2519b23d619eSStefano Zampini AND 2520b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2521b23d619eSStefano Zampini */ 2522b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2523f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2524f4ddd8eeSStefano Zampini } 2525f4ddd8eeSStefano Zampini /* destroy old objects */ 2526f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2527f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2528f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2529a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2530b371cd4fSStefano Zampini n_B = pcis->n_B; 2531b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2532b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 25333a50541eSStefano Zampini 2534a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 25356816873aSStefano Zampini 253653892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2537df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2538854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2539a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2540a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25410e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2542a64d13efSStefano Zampini } 2543a64d13efSStefano Zampini 2544a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 25454641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 25466816873aSStefano Zampini idx_R_local[n_R++] = i; 2547a64d13efSStefano Zampini } 2548a64d13efSStefano Zampini } 2549df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2550df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 25516816873aSStefano Zampini 2552df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2553df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 25546816873aSStefano Zampini } 25553a50541eSStefano Zampini 25563a50541eSStefano Zampini /* Block code */ 25573a50541eSStefano Zampini vbs = 1; 25583a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 25593a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 25603a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 25613a50541eSStefano Zampini PetscInt *vary; 2562df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2563785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 25643a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2565d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2566d3df7717SStefano 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 */ 25670e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2568d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 25693a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 25703a50541eSStefano Zampini is_blocked = PETSC_FALSE; 25713a50541eSStefano Zampini break; 25723a50541eSStefano Zampini } 25733a50541eSStefano Zampini } 2574d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2575d3df7717SStefano Zampini } else { 2576d3df7717SStefano Zampini /* Verify directly the R set */ 2577d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2578d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2579d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2580d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2581d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2582d3df7717SStefano Zampini break; 2583d3df7717SStefano Zampini } 2584d3df7717SStefano Zampini } 2585d3df7717SStefano Zampini } 2586d3df7717SStefano Zampini } 25873a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 25883a50541eSStefano Zampini vbs = bs; 25893a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 25903a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 25913a50541eSStefano Zampini } 25923a50541eSStefano Zampini } 25933a50541eSStefano Zampini } 25943a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2595df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2596df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 259753892102SStefano Zampini 2598df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2599df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 260053892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2601df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 260253892102SStefano Zampini } else { 26033a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 260453892102SStefano Zampini } 2605a64d13efSStefano Zampini 2606a64d13efSStefano Zampini /* print some info if requested */ 2607a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2608a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2609a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26101575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2611a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2612a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 26134f1b2e48SStefano 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); 2614a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2615a64d13efSStefano Zampini } 2616a64d13efSStefano Zampini 2617a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2618df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 26196816873aSStefano Zampini IS is_aux1,is_aux2; 26206816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 26216816873aSStefano Zampini 26223a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2623854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2624854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2625a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26264641a718SStefano Zampini for (i=0; i<n_D; i++) { 26274641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26284641a718SStefano Zampini } 2629a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2630a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26314641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 26324641a718SStefano Zampini aux_array1[j++] = i; 2633a64d13efSStefano Zampini } 2634a64d13efSStefano Zampini } 2635a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2636a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2637a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 26384641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 26394641a718SStefano Zampini aux_array2[j++] = i; 2640a64d13efSStefano Zampini } 2641a64d13efSStefano Zampini } 2642a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2643a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2644a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2645a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2646a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2647a64d13efSStefano Zampini 26488eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2649785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2650a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26514641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 26524641a718SStefano Zampini aux_array1[j++] = i; 2653a64d13efSStefano Zampini } 2654a64d13efSStefano Zampini } 2655a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2656a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2657a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2658a64d13efSStefano Zampini } 26594641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 26603a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2661d62866d3SStefano Zampini } else { 2662df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26636816873aSStefano Zampini IS tis; 26646816873aSStefano Zampini PetscInt schur_size; 26656816873aSStefano Zampini 2666df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 26676816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2668df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 26696816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 26706816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 26716816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 26726816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 26736816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2674d62866d3SStefano Zampini } 2675d62866d3SStefano Zampini } 2676a64d13efSStefano Zampini PetscFunctionReturn(0); 2677a64d13efSStefano Zampini } 2678a64d13efSStefano Zampini 2679304d26faSStefano Zampini 2680304d26faSStefano Zampini #undef __FUNCT__ 2681304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2682684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2683304d26faSStefano Zampini { 2684304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2685304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2686304d26faSStefano Zampini PC pc_temp; 2687304d26faSStefano Zampini Mat A_RR; 2688f4ddd8eeSStefano Zampini MatReuse reuse; 2689304d26faSStefano Zampini PetscScalar m_one = -1.0; 2690304d26faSStefano Zampini PetscReal value; 269104708bb6SStefano Zampini PetscInt n_D,n_R; 26929577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2693304d26faSStefano Zampini PetscErrorCode ierr; 2694e604994aSStefano Zampini /* prefixes stuff */ 2695312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2696e604994aSStefano Zampini size_t len; 2697304d26faSStefano Zampini 2698304d26faSStefano Zampini PetscFunctionBegin; 2699304d26faSStefano Zampini 2700e604994aSStefano Zampini /* compute prefixes */ 2701e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2702e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2703e604994aSStefano Zampini if (!pcbddc->current_level) { 2704e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2705e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2706e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2707e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2708e604994aSStefano Zampini } else { 2709e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2710312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2711e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2712e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2713312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2714312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 271534d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 271634d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2717e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2718e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2719e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2720e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2721e604994aSStefano Zampini } 2722e604994aSStefano Zampini 2723304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2724684f6988SStefano Zampini if (dirichlet) { 2725d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2726450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2727450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2728450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2729450f8f5eSStefano Zampini } 2730450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2731a3df083aSStefano Zampini Mat A_IIn; 2732a3df083aSStefano Zampini 2733a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2734a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2735a3df083aSStefano Zampini pcis->A_II = A_IIn; 2736a3df083aSStefano Zampini } 2737450f8f5eSStefano Zampini } 27383301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27393301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2740964fefecSStefano Zampini } 2741ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2742964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2743304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2744304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2745304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2746304d26faSStefano Zampini /* default */ 2747304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2748e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 27499577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2750304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 27519577ea80SStefano Zampini if (issbaij) { 27529577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 27539577ea80SStefano Zampini } else { 2754304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 27559577ea80SStefano Zampini } 2756304d26faSStefano Zampini /* Allow user's customization */ 2757304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2758304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2759304d26faSStefano Zampini } 2760d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2761df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2762df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2763d62866d3SStefano Zampini 2764df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2765d5574798SStefano Zampini } 2766304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2767304d26faSStefano Zampini if (!n_D) { 2768304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2769304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2770304d26faSStefano Zampini } 2771304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2772304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2773304d26faSStefano Zampini /* set ksp_D into pcis data */ 2774304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2775304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2776304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2777684f6988SStefano Zampini } 2778304d26faSStefano Zampini 2779304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2780684f6988SStefano Zampini A_RR = 0; 2781684f6988SStefano Zampini if (neumann) { 2782d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 278304708bb6SStefano Zampini PetscInt ibs,mbs; 278404708bb6SStefano Zampini PetscBool issbaij; 278504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2786f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 27878ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2788f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2789f4ddd8eeSStefano Zampini PetscInt nn_R; 279081d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2791f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2792f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2793f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2794f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2795f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2796f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2797f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2798727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2799f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2800f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2801f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2802f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2803f4ddd8eeSStefano Zampini } 2804f4ddd8eeSStefano Zampini } 2805f4ddd8eeSStefano Zampini /* last check */ 2806d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2807f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2808f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2809f4ddd8eeSStefano Zampini } 2810f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2811f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2812f4ddd8eeSStefano Zampini } 2813f4ddd8eeSStefano Zampini /* extract A_RR */ 2814af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2815af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 281604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 281704708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 281804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 281904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 282004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2821af732b37SStefano Zampini } else { 282204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 28236816873aSStefano Zampini } 282404708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 282504708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 282604708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 282704708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 282804708bb6SStefano Zampini } else { 282904708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 283004708bb6SStefano Zampini } 283104708bb6SStefano Zampini } 2832f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 28333301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28343301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 28356816873aSStefano Zampini } 2836a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 2837a3df083aSStefano Zampini Mat A_RRn; 2838a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RRn);CHKERRQ(ierr); 2839a3df083aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2840a3df083aSStefano Zampini A_RR = A_RRn; 2841a3df083aSStefano Zampini } 2842f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2843304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2844304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2845304d26faSStefano Zampini /* default */ 2846304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2847e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2848304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 28499577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 28509577ea80SStefano Zampini if (issbaij) { 28519577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28529577ea80SStefano Zampini } else { 2853304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28549577ea80SStefano Zampini } 2855304d26faSStefano Zampini /* Allow user's customization */ 2856304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2857304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2858304d26faSStefano Zampini } 2859304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2860304d26faSStefano Zampini if (!n_R) { 2861304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2862304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2863304d26faSStefano Zampini } 2864df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2865df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2866df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2867d62866d3SStefano Zampini 2868df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2869d62866d3SStefano Zampini } 2870ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2871304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2872304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2873684f6988SStefano Zampini } 28746816873aSStefano Zampini /* free Neumann problem's matrix */ 28756816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2876304d26faSStefano Zampini 2877304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 28780fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2879684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2880684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28811575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2882684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2883684f6988SStefano Zampini } 2884684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 28850fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 28860fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 28870fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 28880fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 28890fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2890304d26faSStefano Zampini /* need to be adapted? */ 2891b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2892b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2893b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2894304d26faSStefano Zampini /* print info */ 2895304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2896e604994aSStefano 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); 2897304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2898304d26faSStefano Zampini } 2899b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2900298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2901304d26faSStefano Zampini } 2902684f6988SStefano Zampini } 2903684f6988SStefano Zampini if (neumann) { /* Neumann */ 29046816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 29050fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 29060fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 29070fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 29080fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 29090fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2910304d26faSStefano Zampini /* need to be adapted? */ 2911b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2912b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2913304d26faSStefano Zampini /* print info */ 2914304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2915e604994aSStefano 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); 2916304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2917304d26faSStefano Zampini } 2918b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2919298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2920304d26faSStefano Zampini } 29210fccc4e9SStefano Zampini } 2922684f6988SStefano Zampini } 2923304d26faSStefano Zampini PetscFunctionReturn(0); 2924304d26faSStefano Zampini } 2925304d26faSStefano Zampini 2926304d26faSStefano Zampini #undef __FUNCT__ 2927ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 292880677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2929674ae819SStefano Zampini { 2930674ae819SStefano Zampini PetscErrorCode ierr; 2931674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2932be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2933674ae819SStefano Zampini 2934674ae819SStefano Zampini PetscFunctionBegin; 2935df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 293680677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 293720c7b377SStefano Zampini } 293880677318SStefano Zampini if (!pcbddc->switch_static) { 293980677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 294080677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 294180677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 294220c7b377SStefano Zampini } 2943df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 294480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 294580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 294620c7b377SStefano Zampini } else { 2947df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2948be83ff47SStefano Zampini 2949df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2950df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 295120c7b377SStefano Zampini } 2952be83ff47SStefano Zampini } else { 295380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 295480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 295580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 295680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 295780677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 295880677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 295980677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 296080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 296180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2962674ae819SStefano Zampini } 2963674ae819SStefano Zampini } 2964df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 296580677318SStefano Zampini if (applytranspose) { 296680677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 296780677318SStefano Zampini } else { 296880677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 296980677318SStefano Zampini } 2970be83ff47SStefano Zampini } else { 2971df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2972be83ff47SStefano Zampini 2973be83ff47SStefano Zampini if (applytranspose) { 2974df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2975be83ff47SStefano Zampini } else { 2976df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2977be83ff47SStefano Zampini } 2978be83ff47SStefano Zampini } 297980677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 298080677318SStefano Zampini if (!pcbddc->switch_static) { 2981df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 298280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2984be83ff47SStefano Zampini } else { 2985df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2986be83ff47SStefano Zampini 2987df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2988df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2989be83ff47SStefano Zampini } 299080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 299180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 299280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 299380677318SStefano Zampini } 299480677318SStefano Zampini } else { 299580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 299980677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 300080677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 300180677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 300280677318SStefano Zampini } 300380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3007674ae819SStefano Zampini } 3008674ae819SStefano Zampini PetscFunctionReturn(0); 3009674ae819SStefano Zampini } 3010674ae819SStefano Zampini 3011dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3012674ae819SStefano Zampini #undef __FUNCT__ 3013674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3014dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3015674ae819SStefano Zampini { 3016674ae819SStefano Zampini PetscErrorCode ierr; 3017674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3018674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3019674ae819SStefano Zampini const PetscScalar zero = 0.0; 3020674ae819SStefano Zampini 3021674ae819SStefano Zampini PetscFunctionBegin; 3022dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 30234fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3024dc359a40SStefano Zampini if (applytranspose) { 3025674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 30268eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3027dc359a40SStefano Zampini } else { 3028674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3029674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 303015aaf578SStefano Zampini } 30314fee134fSStefano Zampini } else { 30324fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 30334fee134fSStefano Zampini } 3034efc2fbd9SStefano Zampini 3035efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 30364f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3037efc2fbd9SStefano Zampini PetscScalar *array; 30384f1b2e48SStefano Zampini PetscInt j; 3039efc2fbd9SStefano Zampini 3040efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30414f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3042efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3043efc2fbd9SStefano Zampini } 3044efc2fbd9SStefano Zampini 304512edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 304612edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 304712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 304812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 304912edc857SStefano Zampini 30509f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 305112edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 305212edc857SStefano Zampini if (pcbddc->coarse_ksp) { 305351694757SStefano Zampini Mat coarse_mat; 3054964fefecSStefano Zampini Vec rhs,sol; 305551694757SStefano Zampini MatNullSpace nullsp; 3056964fefecSStefano Zampini 3057964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3058964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 305951694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 306051694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 306151694757SStefano Zampini if (nullsp) { 306251694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 306351694757SStefano Zampini } 306412edc857SStefano Zampini if (applytranspose) { 3065964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 306612edc857SStefano Zampini } else { 3067964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 306812edc857SStefano Zampini } 306951694757SStefano Zampini if (nullsp) { 307051694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 307151694757SStefano Zampini } 307212edc857SStefano Zampini } 3073674ae819SStefano Zampini 3074674ae819SStefano Zampini /* Local solution on R nodes */ 30754fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 307680677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 30779f00e9b4SStefano Zampini } 30789f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 30799f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 308012edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3081674ae819SStefano Zampini 30824fee134fSStefano Zampini /* Sum contributions from the two levels */ 30834fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3084dc359a40SStefano Zampini if (applytranspose) { 3085dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3086dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3087dc359a40SStefano Zampini } else { 3088674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 30898eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3090dc359a40SStefano Zampini } 3091efc2fbd9SStefano Zampini /* store p0 */ 30924f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3093efc2fbd9SStefano Zampini PetscScalar *array; 30944f1b2e48SStefano Zampini PetscInt j; 3095efc2fbd9SStefano Zampini 3096efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30974f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3098efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3099efc2fbd9SStefano Zampini } 31004fee134fSStefano Zampini } else { /* expand the coarse solution */ 31014fee134fSStefano Zampini if (applytranspose) { 31024fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31034fee134fSStefano Zampini } else { 31044fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31054fee134fSStefano Zampini } 31064fee134fSStefano Zampini } 3107674ae819SStefano Zampini PetscFunctionReturn(0); 3108674ae819SStefano Zampini } 3109674ae819SStefano Zampini 3110674ae819SStefano Zampini #undef __FUNCT__ 3111674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 311212edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3113674ae819SStefano Zampini { 3114674ae819SStefano Zampini PetscErrorCode ierr; 3115674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 311658da7f69SStefano Zampini PetscScalar *array; 311712edc857SStefano Zampini Vec from,to; 3118674ae819SStefano Zampini 3119674ae819SStefano Zampini PetscFunctionBegin; 312012edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 312112edc857SStefano Zampini from = pcbddc->coarse_vec; 312212edc857SStefano Zampini to = pcbddc->vec1_P; 312312edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 312412edc857SStefano Zampini Vec tvec; 312558da7f69SStefano Zampini 312658da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 312758da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 312812edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 312958da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 313058da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 313158da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 313212edc857SStefano Zampini } 313312edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 313412edc857SStefano Zampini from = pcbddc->vec1_P; 313512edc857SStefano Zampini to = pcbddc->coarse_vec; 313612edc857SStefano Zampini } 313712edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3138674ae819SStefano Zampini PetscFunctionReturn(0); 3139674ae819SStefano Zampini } 3140674ae819SStefano Zampini 3141674ae819SStefano Zampini #undef __FUNCT__ 3142674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 314312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3144674ae819SStefano Zampini { 3145674ae819SStefano Zampini PetscErrorCode ierr; 3146674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 314758da7f69SStefano Zampini PetscScalar *array; 314812edc857SStefano Zampini Vec from,to; 3149674ae819SStefano Zampini 3150674ae819SStefano Zampini PetscFunctionBegin; 315112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 315212edc857SStefano Zampini from = pcbddc->coarse_vec; 315312edc857SStefano Zampini to = pcbddc->vec1_P; 315412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 315512edc857SStefano Zampini from = pcbddc->vec1_P; 315612edc857SStefano Zampini to = pcbddc->coarse_vec; 315712edc857SStefano Zampini } 315812edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 315912edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 316012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 316112edc857SStefano Zampini Vec tvec; 316258da7f69SStefano Zampini 316312edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 316458da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 316558da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 316658da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 316758da7f69SStefano Zampini } 316858da7f69SStefano Zampini } else { 316958da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 317058da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 317112edc857SStefano Zampini } 317212edc857SStefano Zampini } 3173674ae819SStefano Zampini PetscFunctionReturn(0); 3174674ae819SStefano Zampini } 3175674ae819SStefano Zampini 3176984c4197SStefano Zampini /* uncomment for testing purposes */ 3177984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3178674ae819SStefano Zampini #undef __FUNCT__ 3179674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3180674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3181674ae819SStefano Zampini { 3182674ae819SStefano Zampini PetscErrorCode ierr; 3183674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3184674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3185674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3186984c4197SStefano Zampini /* one and zero */ 3187984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3188984c4197SStefano Zampini /* space to store constraints and their local indices */ 31899162d606SStefano Zampini PetscScalar *constraints_data; 31909162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 31919162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 31929162d606SStefano Zampini PetscInt *constraints_n; 3193984c4197SStefano Zampini /* iterators */ 3194b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3195984c4197SStefano Zampini /* BLAS integers */ 3196e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3197e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3198c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3199727cdba6SStefano Zampini /* reuse */ 32000e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 32010e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3202984c4197SStefano Zampini /* change of basis */ 3203b3d85658SStefano Zampini PetscBool qr_needed; 32049162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3205984c4197SStefano Zampini /* auxiliary stuff */ 320664efe560SStefano Zampini PetscInt *nnz,*is_indices; 32078a0068c3SStefano Zampini PetscInt ncc; 3208984c4197SStefano Zampini /* some quantities */ 320945a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3210a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3211984c4197SStefano Zampini 3212674ae819SStefano Zampini PetscFunctionBegin; 32138e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 32148e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 32158e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3216088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3217088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 32180e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 32190e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 32200e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32210e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32220e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3223088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3224cf5a6209SStefano Zampini 3225cf5a6209SStefano Zampini /* print some info */ 3226cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3227cf5a6209SStefano Zampini IS vertices; 3228cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3229cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3230cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3231cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 32321575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3233cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3234cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3235fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3236fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3237cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32381575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3239cf5a6209SStefano Zampini } 3240cf5a6209SStefano Zampini 3241cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 32429162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3243cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3244cf5a6209SStefano Zampini const Vec *nearnullvecs; 3245cf5a6209SStefano Zampini Vec *localnearnullsp; 3246cf5a6209SStefano Zampini PetscScalar *array; 3247cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3248cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3249674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3250b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3251674ae819SStefano Zampini PetscScalar *work; 3252674ae819SStefano Zampini PetscReal *singular_vals; 3253674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3254674ae819SStefano Zampini PetscReal *rwork; 3255674ae819SStefano Zampini #endif 3256674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3257674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3258674ae819SStefano Zampini #else 3259964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3260964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3261674ae819SStefano Zampini #endif 3262674ae819SStefano Zampini 3263674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3264d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3265d06fc5fdSStefano Zampini /* free unneeded index sets */ 3266d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3267d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3268674ae819SStefano Zampini } 3269d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3270d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3271d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3272d06fc5fdSStefano Zampini } 3273d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3274d06fc5fdSStefano Zampini n_ISForEdges = 0; 3275d06fc5fdSStefano Zampini } 3276d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3277d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3278d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3279d06fc5fdSStefano Zampini } 3280d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3281d06fc5fdSStefano Zampini n_ISForFaces = 0; 3282d06fc5fdSStefano Zampini } 328370022509SStefano Zampini 328470022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 328570022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 328670022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 328770022509SStefano Zampini if (pcbddc->NullSpace) { 328870022509SStefano Zampini PetscBool tbool[2],gbool[2]; 328970022509SStefano Zampini 329070022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3291b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3292d06fc5fdSStefano Zampini if (!ISForEdges) { 3293d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3294d06fc5fdSStefano Zampini } 3295b8ffe317SStefano Zampini } 3296d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3297d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3298d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3299d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3300d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 330198a51de6SStefano Zampini } 330270022509SStefano Zampini #endif 330308122e43SStefano Zampini 3304674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3305674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3306674ae819SStefano Zampini if (nearnullsp) { 3307674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3308f4ddd8eeSStefano Zampini /* remove any stored info */ 3309f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3310f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3311f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3312f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3313f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3314473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3315f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3316f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3317f4ddd8eeSStefano Zampini } 3318984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3319984c4197SStefano Zampini nnsp_size = 0; 3320674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3321674ae819SStefano Zampini } 3322984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3323984c4197SStefano Zampini max_constraints = nnsp_size; 3324984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3325984c4197SStefano Zampini 3326674ae819SStefano Zampini /* 3327674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 33289162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 33299162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 33309162d606SStefano Zampini There can be multiple constraints per connected component 3331674ae819SStefano Zampini */ 3332674ae819SStefano Zampini n_vertices = 0; 3333674ae819SStefano Zampini if (ISForVertices) { 3334674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3335674ae819SStefano Zampini } 33369162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 33379162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 33389162d606SStefano Zampini 33399162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 33409162d606SStefano Zampini total_counts *= max_constraints; 3341674ae819SStefano Zampini total_counts += n_vertices; 33424641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 33439162d606SStefano Zampini 3344674ae819SStefano Zampini total_counts = 0; 3345674ae819SStefano Zampini max_size_of_constraint = 0; 3346674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 33479162d606SStefano Zampini IS used_is; 3348674ae819SStefano Zampini if (i<n_ISForEdges) { 33499162d606SStefano Zampini used_is = ISForEdges[i]; 3350674ae819SStefano Zampini } else { 33519162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3352674ae819SStefano Zampini } 33539162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3354674ae819SStefano Zampini total_counts += j; 3355674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3356674ae819SStefano Zampini } 33579162d606SStefano 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); 33589162d606SStefano Zampini 3359984c4197SStefano Zampini /* get local part of global near null space vectors */ 3360785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3361984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3362984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3363e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3364e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3365984c4197SStefano Zampini } 3366674ae819SStefano Zampini 3367242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3368242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3369a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3370242a89d7SStefano Zampini 3371984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3372a773dcb8SStefano Zampini if (!skip_lapack) { 3373674ae819SStefano Zampini PetscScalar temp_work; 3374911cabfeSStefano Zampini 3375674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3376984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3377785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3378785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3379785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3380674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3381785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3382674ae819SStefano Zampini #endif 3383674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3384c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3385c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3386674ae819SStefano Zampini lwork = -1; 3387674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3388674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3389c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3390674ae819SStefano Zampini #else 3391c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3392674ae819SStefano Zampini #endif 3393674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3394984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3395674ae819SStefano Zampini #else /* on missing GESVD */ 3396674ae819SStefano Zampini /* SVD */ 3397674ae819SStefano Zampini PetscInt max_n,min_n; 3398674ae819SStefano Zampini max_n = max_size_of_constraint; 3399984c4197SStefano Zampini min_n = max_constraints; 3400984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3401674ae819SStefano Zampini min_n = max_size_of_constraint; 3402984c4197SStefano Zampini max_n = max_constraints; 3403674ae819SStefano Zampini } 3404785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3405674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3406785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3407674ae819SStefano Zampini #endif 3408674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3409674ae819SStefano Zampini lwork = -1; 3410e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3411e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3412b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3413674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3414674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 34159162d606SStefano 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)); 3416674ae819SStefano Zampini #else 34179162d606SStefano 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)); 3418674ae819SStefano Zampini #endif 3419674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3420984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3421984c4197SStefano Zampini #endif /* on missing GESVD */ 3422674ae819SStefano Zampini /* Allocate optimal workspace */ 3423674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3424854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3425674ae819SStefano Zampini } 3426674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3427674ae819SStefano Zampini total_counts = 0; 34289162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 34299162d606SStefano Zampini constraints_data_ptr[0] = 0; 3430674ae819SStefano Zampini /* vertices */ 34319162d606SStefano Zampini if (n_vertices) { 3432674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 34339162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3434674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 34359162d606SStefano Zampini constraints_n[total_counts] = 1; 34369162d606SStefano Zampini constraints_data[total_counts] = 1.0; 34379162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 34389162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3439674ae819SStefano Zampini total_counts++; 3440674ae819SStefano Zampini } 3441674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3442674ae819SStefano Zampini n_vertices = total_counts; 3443674ae819SStefano Zampini } 3444984c4197SStefano Zampini 3445674ae819SStefano Zampini /* edges and faces */ 34469162d606SStefano Zampini total_counts_cc = total_counts; 3447911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 34489162d606SStefano Zampini IS used_is; 34499162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 34509162d606SStefano Zampini 3451911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 34529162d606SStefano Zampini used_is = ISForEdges[ncc]; 3453984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3454674ae819SStefano Zampini } else { 34559162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3456984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3457674ae819SStefano Zampini } 3458674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 34599162d606SStefano Zampini 34609162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 34619162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3462984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3463984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3464674ae819SStefano Zampini if (nnsp_has_cnst) { 34655b08dc53SStefano Zampini PetscScalar quad_value; 34669162d606SStefano Zampini 34679162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34689162d606SStefano Zampini idxs_copied = PETSC_TRUE; 34699162d606SStefano Zampini 3470a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3471674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3472a773dcb8SStefano Zampini } else { 3473a773dcb8SStefano Zampini quad_value = 1.0; 3474a773dcb8SStefano Zampini } 3475674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34769162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3477674ae819SStefano Zampini } 34789162d606SStefano Zampini temp_constraints++; 3479674ae819SStefano Zampini total_counts++; 3480674ae819SStefano Zampini } 3481674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3482984c4197SStefano Zampini PetscReal real_value; 34839162d606SStefano Zampini PetscScalar *ptr_to_data; 34849162d606SStefano Zampini 3485984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 34869162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3487674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34889162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3489674ae819SStefano Zampini } 3490984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3491984c4197SStefano Zampini /* check if array is null on the connected component */ 3492e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34939162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 34945b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3495674ae819SStefano Zampini temp_constraints++; 3496674ae819SStefano Zampini total_counts++; 34979162d606SStefano Zampini if (!idxs_copied) { 34989162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34999162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3500674ae819SStefano Zampini } 3501674ae819SStefano Zampini } 35029162d606SStefano Zampini } 35039162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 350445a1bb75SStefano Zampini valid_constraints = temp_constraints; 3505eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3506a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 35079162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 35089162d606SStefano Zampini 35099162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3510a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35119162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3512a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 35139162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3514a773dcb8SStefano Zampini } else { /* perform SVD */ 3515984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 35169162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3517674ae819SStefano Zampini 3518674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3519984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3520984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3521984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3522984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3523984c4197SStefano Zampini from that computed using LAPACKgesvd 3524984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3525984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3526984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3527674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3528e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3529984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3530674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3531674ae819SStefano Zampini for (k=0;k<j+1;k++) { 35329162d606SStefano 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)); 3533674ae819SStefano Zampini } 3534674ae819SStefano Zampini } 3535e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3536e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3537e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3538674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3539c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3540674ae819SStefano Zampini #else 3541c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3542674ae819SStefano Zampini #endif 3543674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3544984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3545984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3546674ae819SStefano Zampini j = 0; 3547984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3548674ae819SStefano Zampini total_counts = total_counts-j; 354945a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3550e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3551c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3552c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3553c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3554c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3555c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3556c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3557674ae819SStefano Zampini if (j<temp_constraints) { 3558984c4197SStefano Zampini PetscInt ii; 3559984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3560674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 35619162d606SStefano 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)); 3562674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3563984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3564674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 35659162d606SStefano 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]; 3566674ae819SStefano Zampini } 3567674ae819SStefano Zampini } 3568674ae819SStefano Zampini } 3569674ae819SStefano Zampini #else /* on missing GESVD */ 3570e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3571e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3572b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3573674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3574674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35759162d606SStefano 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)); 3576674ae819SStefano Zampini #else 35779162d606SStefano 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)); 3578674ae819SStefano Zampini #endif 3579984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3580674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3581984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3582e310c8b4SStefano Zampini k = temp_constraints; 3583e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3584674ae819SStefano Zampini j = 0; 3585e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 358645a1bb75SStefano Zampini valid_constraints = k-j; 3587911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3588984c4197SStefano Zampini #endif /* on missing GESVD */ 3589674ae819SStefano Zampini } 3590a773dcb8SStefano Zampini } 35919162d606SStefano Zampini /* update pointers information */ 35929162d606SStefano Zampini if (valid_constraints) { 35939162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 35949162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 35959162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 35969162d606SStefano Zampini /* set change_of_basis flag */ 359745a1bb75SStefano Zampini if (boolforchange) { 3598b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 35999162d606SStefano Zampini } 3600b3d85658SStefano Zampini total_counts_cc++; 360145a1bb75SStefano Zampini } 360245a1bb75SStefano Zampini } 3603984c4197SStefano Zampini /* free workspace */ 36048f1c130eSStefano Zampini if (!skip_lapack) { 3605984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3606984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3607984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3608984c4197SStefano Zampini #endif 3609984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3610984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3611984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3612984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3613984c4197SStefano Zampini #endif 3614984c4197SStefano Zampini } 3615984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3616984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3617984c4197SStefano Zampini } 3618984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3619cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3620cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3621cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3622cf5a6209SStefano Zampini } 3623cf5a6209SStefano Zampini if (n_ISForFaces) { 3624cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3625cf5a6209SStefano Zampini } 3626cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3627cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3628cf5a6209SStefano Zampini } 3629cf5a6209SStefano Zampini if (n_ISForEdges) { 3630cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3631cf5a6209SStefano Zampini } 3632cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 363308122e43SStefano Zampini } else { 363408122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3635984c4197SStefano Zampini 363608122e43SStefano Zampini total_counts = 0; 363708122e43SStefano Zampini n_vertices = 0; 3638d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3639d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 364008122e43SStefano Zampini } 364108122e43SStefano Zampini max_constraints = 0; 36429162d606SStefano Zampini total_counts_cc = 0; 364308122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 364408122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 36459162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 364608122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 364708122e43SStefano Zampini } 36489162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 36499162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 36509162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 36519162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 365274d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 36539162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 36549162d606SStefano Zampini total_counts_cc = 0; 36559162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 36569162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 36579162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 365808122e43SStefano Zampini } 365908122e43SStefano Zampini } 36609162d606SStefano Zampini #if 0 36619162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 36629162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36639162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 36649162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 36659162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 36669162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 36679162d606SStefano Zampini } 36689162d606SStefano Zampini printf("\n"); 36699162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 36709162d606SStefano Zampini } 36711b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 36728bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 36731b968477SStefano Zampini } 36741b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 36758bec7fa6SStefano 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]); 36761b968477SStefano Zampini } 367708122e43SStefano Zampini #endif 367808122e43SStefano Zampini 36798bec7fa6SStefano Zampini max_size_of_constraint = 0; 36809162d606SStefano 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]); 36819162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 368208122e43SStefano Zampini /* Change of basis */ 3683b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 368408122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 368508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 368608122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3687b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 368808122e43SStefano Zampini } 368908122e43SStefano Zampini } 369008122e43SStefano Zampini } 369108122e43SStefano Zampini } 3692984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 36934f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 369408122e43SStefano Zampini 36959162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 36969162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 36979162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 36989162d606SStefano 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); 369908122e43SStefano Zampini } 3700674ae819SStefano Zampini 3701674ae819SStefano Zampini /* Create constraint matrix */ 3702674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 370316f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3704984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3705984c4197SStefano Zampini 3706984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3707a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3708a717540cSStefano Zampini qr_needed = PETSC_FALSE; 370974d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3710984c4197SStefano Zampini total_primal_vertices=0; 3711b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 37129162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37139162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3714984c4197SStefano Zampini if (size_of_constraint == 1) { 37159162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3716b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 371764efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 37189162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37199162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3720a717540cSStefano Zampini } 3721b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 372274d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3723a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3724a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3725a717540cSStefano Zampini } 3726fa434743SStefano Zampini } else { 3727b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3728fa434743SStefano Zampini } 3729a717540cSStefano Zampini } 3730b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3731b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3732674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 373370022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3734b3d85658SStefano Zampini 37354f1b2e48SStefano 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); 37360e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 37370e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3738984c4197SStefano Zampini 3739984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 374074d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3741785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3742984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 374374d5cdf7SStefano Zampini 3744984c4197SStefano Zampini j = total_primal_vertices; 374574d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3746b3d85658SStefano Zampini cum = total_primal_vertices; 37479162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37484641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3749b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3750b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3751b3d85658SStefano Zampini cum++; 37529162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 375374d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 375474d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 375574d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 375674d5cdf7SStefano Zampini } 37579162d606SStefano Zampini j += constraints_n[i]; 3758674ae819SStefano Zampini } 3759674ae819SStefano Zampini } 3760674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3761674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3762088faed8SStefano Zampini 3763674ae819SStefano Zampini /* set values in constraint matrix */ 3764984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 37650e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3766674ae819SStefano Zampini } 3767984c4197SStefano Zampini total_counts = total_primal_vertices; 37689162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37694641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 37709162d606SStefano Zampini PetscInt *cols; 37719162d606SStefano Zampini 37729162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 37739162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 37749162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37759162d606SStefano Zampini PetscInt row = total_counts+k; 37769162d606SStefano Zampini PetscScalar *vals; 37779162d606SStefano Zampini 37789162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 37799162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 37809162d606SStefano Zampini } 37819162d606SStefano Zampini total_counts += constraints_n[i]; 3782674ae819SStefano Zampini } 3783674ae819SStefano Zampini } 3784674ae819SStefano Zampini /* assembling */ 3785674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3786674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3787088faed8SStefano Zampini 3788984c4197SStefano Zampini /* 378945a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3790984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3791984c4197SStefano Zampini */ 3792674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3793674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3794026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3795984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3796984c4197SStefano Zampini /* working stuff for GEQRF */ 379781d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3798984c4197SStefano Zampini PetscBLASInt lqr_work; 3799984c4197SStefano Zampini /* working stuff for UNGQR */ 3800984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3801984c4197SStefano Zampini PetscBLASInt lgqr_work; 3802984c4197SStefano Zampini /* working stuff for TRTRS */ 3803984c4197SStefano Zampini PetscScalar *trs_rhs; 38043f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3805984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3806984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3807984c4197SStefano Zampini PetscScalar *start_vals; 3808984c4197SStefano Zampini /* working stuff for values insertion */ 38094641a718SStefano Zampini PetscBT is_primal; 381064efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3811906d46d4SStefano Zampini /* matrix sizes */ 3812906d46d4SStefano Zampini PetscInt global_size,local_size; 3813906d46d4SStefano Zampini /* temporary change of basis */ 3814906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3815cf5a6209SStefano Zampini /* extra space for debugging */ 3816cf5a6209SStefano Zampini PetscScalar *dbg_work; 3817984c4197SStefano Zampini 3818906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3819906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 382016f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3821bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3822906d46d4SStefano Zampini /* nonzeros for local mat */ 3823bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3824bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 38259162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3826a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 38279162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3828a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 38299162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3830a717540cSStefano Zampini } else { 38319162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 38329162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3833a717540cSStefano Zampini } 3834a717540cSStefano Zampini } 3835a717540cSStefano Zampini } 3836906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3837bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3838a717540cSStefano Zampini /* Set initial identity in the matrix */ 3839bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3840906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3841a717540cSStefano Zampini } 3842a717540cSStefano Zampini 3843a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3844a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3845a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3846a717540cSStefano Zampini } 3847a717540cSStefano Zampini 3848a717540cSStefano Zampini 3849a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3850a717540cSStefano Zampini /* 3851a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3852a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3853a717540cSStefano Zampini 3854a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3855a717540cSStefano Zampini 3856a6b551f4SStefano 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) 3857a6b551f4SStefano Zampini 3858a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3859a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3860a717540cSStefano Zampini | ... | 3861a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3862a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3863a717540cSStefano Zampini 3864a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3865a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3866a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3867a6b551f4SStefano Zampini 3868a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3869a717540cSStefano Zampini */ 3870a717540cSStefano Zampini if (qr_needed) { 3871984c4197SStefano Zampini /* space to store Q */ 3872854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3873984c4197SStefano Zampini /* first we issue queries for optimal work */ 38743f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38753f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 38763f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3877984c4197SStefano Zampini lqr_work = -1; 38783f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3879984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3880984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3881785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3882984c4197SStefano Zampini lgqr_work = -1; 38833f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38843f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 38853f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 38863f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38873f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 38883f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3889984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3890984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3891785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3892984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3893785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3894984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3895785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3896a717540cSStefano Zampini /* allocating workspace for check */ 3897a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3898cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3899a717540cSStefano Zampini } 3900a717540cSStefano Zampini } 3901984c4197SStefano Zampini /* array to store whether a node is primal or not */ 39024641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3903473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 39040e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 390539e2fb2aSStefano Zampini if (i != total_primal_vertices) { 390639e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 39074641a718SStefano Zampini } 390839e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 390939e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 391039e2fb2aSStefano Zampini } 391139e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3912984c4197SStefano Zampini 3913a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 39149162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 39159162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 39164641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3917984c4197SStefano Zampini /* get constraint info */ 39189162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3919984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3920984c4197SStefano Zampini 3921984c4197SStefano Zampini if (pcbddc->dbg_flag) { 39229162d606SStefano 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); 3923674ae819SStefano Zampini } 3924984c4197SStefano Zampini 3925fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3926a717540cSStefano Zampini 3927a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3928a717540cSStefano Zampini if (pcbddc->dbg_flag) { 39299162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3930a717540cSStefano Zampini } 3931984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 39329162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3933984c4197SStefano Zampini 3934984c4197SStefano Zampini /* compute QR decomposition of constraints */ 39353f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39363f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3938674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39393f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3940984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3941674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3942984c4197SStefano Zampini 3943984c4197SStefano Zampini /* explictly compute R^-T */ 3944984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3945984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 39463f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39473f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 39483f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39493f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3950984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39513f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3952984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3953984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3954984c4197SStefano Zampini 3955a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 39563f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39573f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39583f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39593f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3960984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39613f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3962984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3963984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3964984c4197SStefano Zampini 3965984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3966984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3967984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 39683f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39693f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39703f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39713f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39723f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 39733f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3974984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39759162d606SStefano 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)); 3976984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39779162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3978984c4197SStefano Zampini 3979984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 39809162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3981984c4197SStefano Zampini /* insert cols for primal dofs */ 3982984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3983984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 39849162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3985906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3986984c4197SStefano Zampini } 3987984c4197SStefano Zampini /* insert cols for dual dofs */ 3988984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 39899162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3990984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 39919162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3992906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3993984c4197SStefano Zampini j++; 3994674ae819SStefano Zampini } 3995674ae819SStefano Zampini } 3996984c4197SStefano Zampini 3997984c4197SStefano Zampini /* check change of basis */ 3998984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3999984c4197SStefano Zampini PetscInt ii,jj; 4000984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4001c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4002c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4003c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4004c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4005c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4006c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4007984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4008cf5a6209SStefano 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)); 4009984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4010984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4011984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4012cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4013cf5a6209SStefano 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; 4014674ae819SStefano Zampini } 4015674ae819SStefano Zampini } 4016984c4197SStefano Zampini if (!valid_qr) { 401722d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4018984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4019984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4020cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4021cf5a6209SStefano 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])); 4022674ae819SStefano Zampini } 4023cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4024cf5a6209SStefano 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])); 4025984c4197SStefano Zampini } 4026984c4197SStefano Zampini } 4027984c4197SStefano Zampini } 4028674ae819SStefano Zampini } else { 402922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4030674ae819SStefano Zampini } 4031674ae819SStefano Zampini } 4032a717540cSStefano Zampini } else { /* simple transformation block */ 4033a717540cSStefano Zampini PetscInt row,col; 4034a6b551f4SStefano Zampini PetscScalar val,norm; 4035a6b551f4SStefano Zampini 4036a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40379162d606SStefano 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)); 4038a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 40399162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 40409162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4041bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 40429162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4043906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 40449162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4045a717540cSStefano Zampini } else { 4046a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 40479162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4048a717540cSStefano Zampini if (row != col) { 40499162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4050a717540cSStefano Zampini } else { 40519162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4052a717540cSStefano Zampini } 4053906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4054a717540cSStefano Zampini } 4055a717540cSStefano Zampini } 4056a717540cSStefano Zampini } 405798a51de6SStefano Zampini if (pcbddc->dbg_flag) { 405822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4059a717540cSStefano Zampini } 4060674ae819SStefano Zampini } 4061984c4197SStefano Zampini } else { 4062984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40639162d606SStefano 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); 4064674ae819SStefano Zampini } 4065674ae819SStefano Zampini } 4066674ae819SStefano Zampini } 4067a717540cSStefano Zampini 4068a717540cSStefano Zampini /* free workspace */ 4069a717540cSStefano Zampini if (qr_needed) { 4070984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4071cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4072984c4197SStefano Zampini } 4073984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4074984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4075984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4076984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4077984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4078674ae819SStefano Zampini } 4079a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4080906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4081906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4082906d46d4SStefano Zampini 4083906d46d4SStefano Zampini /* assembling of global change of variable */ 408488c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4085bbb9e6c6SStefano Zampini Mat tmat; 408616f15bc4SStefano Zampini PetscInt bs; 408716f15bc4SStefano Zampini 4088906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4089906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4090bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4091bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4092bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4093bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 409416f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 409516f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4096906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4097bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4098bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4099bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4100bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4101bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4102e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4103e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4104bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4105bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 410688c03ad3SStefano Zampini 4107906d46d4SStefano Zampini /* check */ 4108906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4109906d46d4SStefano Zampini PetscReal error; 4110906d46d4SStefano Zampini Vec x,x_change; 4111906d46d4SStefano Zampini 4112906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4113906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4114906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4115906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4116e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4117e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4118bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4119e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4120e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4121906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4122906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4123906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4124906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4125bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4126906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4127906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4128906d46d4SStefano Zampini } 412988c03ad3SStefano Zampini } 4130b96c3477SStefano Zampini 4131b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4132b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4133b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4134b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4135ac632422SStefano Zampini Mat S_new,tmat; 4136b087196eSStefano Zampini ISLocalToGlobalMapping NtoSall; 4137b087196eSStefano Zampini IS is_all_N,is_V,is_V_Sall; 4138b087196eSStefano Zampini const PetscScalar *array; 4139b087196eSStefano Zampini const PetscInt *idxs_V,*idxs_all; 4140b087196eSStefano Zampini PetscInt i,n_V; 4141bbb9e6c6SStefano Zampini 4142bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 41436816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4144b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4145b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4146b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4147b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4148bbb9e6c6SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4149b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4150ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4151b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4152ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4153b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4154b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4155b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4156b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4157b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4158b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4159b087196eSStefano Zampini PetscScalar val; 4160b087196eSStefano Zampini PetscInt idx; 4161b087196eSStefano Zampini 4162b087196eSStefano Zampini idx = idxs_V[i]; 4163b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4164b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4165b087196eSStefano Zampini } 4166b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4167b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4168ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4169ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4170ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4171ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4172b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4173ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4174b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4175ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4176ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4177ac632422SStefano Zampini } 4178b087196eSStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4179b087196eSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4180b087196eSStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 418188c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4182b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 418388c03ad3SStefano Zampini } else { 418488c03ad3SStefano Zampini ierr = ISDestroy(&sub_schurs->change_primal);CHKERRQ(ierr); 418588c03ad3SStefano Zampini sub_schurs->change_primal = is_V_Sall; 418688c03ad3SStefano Zampini ierr = KSPDestroy(&sub_schurs->change);CHKERRQ(ierr); 418788c03ad3SStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&sub_schurs->change);CHKERRQ(ierr); 418888c03ad3SStefano Zampini ierr = KSPSetOperators(sub_schurs->change,tmat,tmat);CHKERRQ(ierr); 418988c03ad3SStefano Zampini } 419088c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4191b96c3477SStefano Zampini } 4192b96c3477SStefano Zampini } 4193906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 4194906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4195b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4196b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4197b9b85e73SStefano Zampini } 4198906d46d4SStefano Zampini 4199906d46d4SStefano Zampini /* set up change of basis context */ 4200906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4201906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4202906d46d4SStefano Zampini 4203906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4204906d46d4SStefano Zampini PetscInt global_size,local_size; 4205906d46d4SStefano Zampini 4206906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4207906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4208906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4209906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4210906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4211906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4212906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4213906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4214906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4215906d46d4SStefano Zampini } else { 4216906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4217906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4218906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4219906d46d4SStefano Zampini } 4220906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4221906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4222906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4223906d46d4SStefano Zampini } else { 4224906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4225906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4226906d46d4SStefano Zampini } 4227906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4228906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4229906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4230906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4231a8661815SStefano Zampini } else { 4232a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4233b9b85e73SStefano Zampini } 4234a717540cSStefano Zampini 42354f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 42364f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 42374f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 42384f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4239019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4240019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4241019a44ceSStefano Zampini pcbddc->local_primal_size++; 4242019a44ceSStefano Zampini } 4243019a44ceSStefano Zampini 4244019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4245727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4246727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 42479f47a83aSStefano 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); 4248c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 42490e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 42509f47a83aSStefano 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); 4251727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4252727cdba6SStefano Zampini } 42530e6343abSStefano Zampini } 42540e6343abSStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4255727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4256727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 4257727cdba6SStefano Zampini 4258a717540cSStefano Zampini /* flush dbg viewer */ 4259b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4260b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4261b8ffe317SStefano Zampini } 4262a717540cSStefano Zampini 4263e310c8b4SStefano Zampini /* free workspace */ 4264a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 42654641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 426608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 42679162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 42689162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 426908122e43SStefano Zampini } else { 42709162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 42719162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 42729162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 427308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 427408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 42759162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 42769162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 427708122e43SStefano Zampini } 4278674ae819SStefano Zampini PetscFunctionReturn(0); 4279674ae819SStefano Zampini } 4280674ae819SStefano Zampini 4281674ae819SStefano Zampini #undef __FUNCT__ 4282674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4283674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4284674ae819SStefano Zampini { 4285674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4286674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4287674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 42887fb0e2dbSStefano Zampini PetscInt ierr,i,vertex_size,N; 4289674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4290674ae819SStefano Zampini 4291674ae819SStefano Zampini PetscFunctionBegin; 42928e61c736SStefano Zampini /* Reset previously computed graph */ 42938e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4294674ae819SStefano Zampini /* Init local Graph struct */ 42957fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 42963bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4297674ae819SStefano Zampini 4298575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 42995099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 43005099eff2SStefano 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); 4301575ad6abSStefano Zampini } 43029577ea80SStefano Zampini 4303674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4304d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 43054d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 43064d379d7bSStefano Zampini PetscInt nvtxs; 4307e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4308674ae819SStefano Zampini 43092fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 43102fffb893SStefano Zampini if (flg_row) { 43114d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4312b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 43132fffb893SStefano Zampini } 43142fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4315674ae819SStefano Zampini } 43169b28b941SStefano Zampini if (pcbddc->dbg_flag) { 43179b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4318674ae819SStefano Zampini } 4319674ae819SStefano Zampini 432063602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 4321674ae819SStefano Zampini vertex_size = 1; 432263602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 432363602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 432495ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 432563602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4326e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 432763602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4328674ae819SStefano Zampini } 432963602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 433063602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 433163602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4332674ae819SStefano Zampini } 433363602bcaSStefano Zampini /* mat block size as vertex size (used for elasticity with rigid body modes as nearnullspace) */ 4334674ae819SStefano Zampini ierr = MatGetBlockSize(matis->A,&vertex_size);CHKERRQ(ierr); 433563602bcaSStefano Zampini } else { 433663602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 433763602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4338854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 433963602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 434063602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 434163602bcaSStefano Zampini } 434263602bcaSStefano Zampini } 4343674ae819SStefano Zampini } 4344674ae819SStefano Zampini 4345674ae819SStefano Zampini /* Setup of Graph */ 4346785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4347e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4348785d1243SStefano Zampini } 4349785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4350e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4351785d1243SStefano Zampini } 435230368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 435330368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 435430368db7SStefano Zampini } 435530368db7SStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4356674ae819SStefano Zampini 43574f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 43584f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 43594f1b2e48SStefano Zampini PetscInt *local_subs; 43604f1b2e48SStefano Zampini 43614f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 43624f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 43634f1b2e48SStefano Zampini const PetscInt *idxs; 43644f1b2e48SStefano Zampini PetscInt nl,j; 43654f1b2e48SStefano Zampini 43664f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 43674f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43684f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 43694f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 43704f1b2e48SStefano Zampini } 43714f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43724f1b2e48SStefano Zampini } 43734f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 43744f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 43754f1b2e48SStefano Zampini } 43764f1b2e48SStefano Zampini 4377674ae819SStefano Zampini /* Graph's connected components analysis */ 4378674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4379674ae819SStefano Zampini 4380674ae819SStefano Zampini /* print some info to stdout */ 4381674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4382302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4383674ae819SStefano Zampini } 4384fb180af4SStefano Zampini 4385fb180af4SStefano Zampini /* mark topography has done */ 4386fb180af4SStefano Zampini pcbddc->recompute_topography = PETSC_FALSE; 4387674ae819SStefano Zampini PetscFunctionReturn(0); 4388674ae819SStefano Zampini } 4389674ae819SStefano Zampini 4390dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4391674ae819SStefano Zampini #undef __FUNCT__ 4392674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4393dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4394674ae819SStefano Zampini { 4395dc456d91SStefano Zampini PetscSF sf; 4396dc456d91SStefano Zampini PetscLayout map; 4397dc456d91SStefano Zampini const PetscInt *idxs; 4398dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4399dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4400dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4401dc456d91SStefano Zampini PetscMPIInt commsize; 4402674ae819SStefano Zampini PetscBool first_found; 4403674ae819SStefano Zampini PetscErrorCode ierr; 4404674ae819SStefano Zampini 4405674ae819SStefano Zampini PetscFunctionBegin; 4406dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4407dc456d91SStefano Zampini if (subset_mult) { 4408dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4409dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4410dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4411674ae819SStefano Zampini } 4412dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4413dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4414dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4415dc456d91SStefano Zampini for (i=0;i<n;i++) { 4416dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4417dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4418674ae819SStefano Zampini } 4419dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4420dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4421dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4422dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4423dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4424dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4425dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4426dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4427dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4428dc456d91SStefano Zampini 4429dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4430dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4431dc456d91SStefano Zampini if (subset_mult) { 4432dc456d91SStefano Zampini const PetscInt* idxs_mult; 4433dc456d91SStefano Zampini 4434dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4435dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4436dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4437674ae819SStefano Zampini } else { 4438dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4439674ae819SStefano Zampini } 4440dc456d91SStefano Zampini /* local size of new subset */ 4441dc456d91SStefano Zampini n_n = 0; 4442dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4443dc456d91SStefano Zampini 4444dc456d91SStefano Zampini /* global indexes in layout */ 4445dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4446dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4447dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4448dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4449dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4450dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4451dc456d91SStefano Zampini 4452dc456d91SStefano Zampini /* reduce from leaves to roots */ 4453dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 445464a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 445564a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4456dc456d91SStefano Zampini 4457dc456d91SStefano Zampini /* count indexes in local part of layout */ 4458674ae819SStefano Zampini nlocals = 0; 4459674ae819SStefano Zampini first_index = -1; 4460674ae819SStefano Zampini first_found = PETSC_FALSE; 4461dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4462dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4463674ae819SStefano Zampini first_found = PETSC_TRUE; 4464674ae819SStefano Zampini first_index = i; 4465674ae819SStefano Zampini } 4466dc456d91SStefano Zampini nlocals += root_data[i]; 4467674ae819SStefano Zampini } 4468dc456d91SStefano Zampini 4469dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 44705fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4471dc456d91SStefano Zampini start = 0; 447264a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44735fa240b1SStefano Zampini #else 447464a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44755fa240b1SStefano Zampini start = start-nlocals; 44765fa240b1SStefano Zampini #endif 44775fa240b1SStefano Zampini 4478dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4479dc456d91SStefano Zampini *N_n = start + nlocals; 4480dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4481dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4482674ae819SStefano Zampini } 44835fa240b1SStefano Zampini 44845fa240b1SStefano Zampini /* adapt root data with cumulative */ 4485674ae819SStefano Zampini if (first_found) { 4486dc456d91SStefano Zampini PetscInt old_index; 4487dc456d91SStefano Zampini 4488dc456d91SStefano Zampini root_data[first_index] += start; 4489674ae819SStefano Zampini old_index = first_index; 4490dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4491dc456d91SStefano Zampini if (root_data[i]) { 4492dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4493674ae819SStefano Zampini old_index = i; 4494674ae819SStefano Zampini } 4495674ae819SStefano Zampini } 4496674ae819SStefano Zampini } 4497dc456d91SStefano Zampini 4498dc456d91SStefano Zampini /* from roots to leaves */ 4499dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4500dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4501dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4502dc456d91SStefano Zampini 4503dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4504dc456d91SStefano Zampini if (subset_mult) { 4505dc456d91SStefano Zampini const PetscInt* idxs_mult; 4506dc456d91SStefano Zampini PetscInt cum; 4507dc456d91SStefano Zampini 4508dc456d91SStefano Zampini cum = 0; 4509dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4510dc456d91SStefano Zampini for (i=0;i<n;i++) { 4511dc456d91SStefano Zampini PetscInt j; 4512dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4513674ae819SStefano Zampini } 4514dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4515674ae819SStefano Zampini } else { 4516dc456d91SStefano Zampini for (i=0;i<n;i++) { 4517dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4518674ae819SStefano Zampini } 4519674ae819SStefano Zampini } 4520dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4521dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4522674ae819SStefano Zampini PetscFunctionReturn(0); 4523674ae819SStefano Zampini } 45249a7d3425SStefano Zampini 45259a7d3425SStefano Zampini #undef __FUNCT__ 45269a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 45279a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 45289a7d3425SStefano Zampini { 45299a7d3425SStefano Zampini PetscInt i,j; 45309a7d3425SStefano Zampini PetscScalar *alphas; 45319a7d3425SStefano Zampini PetscErrorCode ierr; 45329a7d3425SStefano Zampini 45339a7d3425SStefano Zampini PetscFunctionBegin; 45349a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4535785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 45369a7d3425SStefano Zampini for (i=0;i<n;i++) { 45379a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 45389a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 45399a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 45409a7d3425SStefano Zampini } 45419a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 45429a7d3425SStefano Zampini PetscFunctionReturn(0); 45439a7d3425SStefano Zampini } 45449a7d3425SStefano Zampini 4545e7931f94SStefano Zampini #undef __FUNCT__ 454670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4547b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4548e7931f94SStefano Zampini { 454952e5ac9dSStefano Zampini IS ranks_send_to; 4550e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4551e7931f94SStefano Zampini PetscMPIInt size,rank,color; 455252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 455352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 45543837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 45552b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4556e7931f94SStefano Zampini PetscSubcomm subcomm; 455752e5ac9dSStefano Zampini PetscErrorCode ierr; 4558a57a6d2fSStefano Zampini 4559e7931f94SStefano Zampini PetscFunctionBegin; 45602b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 45612b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 45622b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4563e7931f94SStefano Zampini 4564e7931f94SStefano Zampini /* Get info on mapping */ 45653bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 45663bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4567e7931f94SStefano Zampini 4568e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4569785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4570e7931f94SStefano Zampini xadj[0] = 0; 4571e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4572785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4573785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4574e7931f94SStefano Zampini 45752b510759SStefano Zampini if (threshold) { 4576d023bfaeSStefano Zampini PetscInt xadj_count = 0; 45772b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4578d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4579d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4580d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4581d023bfaeSStefano Zampini xadj_count++; 4582e7931f94SStefano Zampini } 4583e7931f94SStefano Zampini } 4584d023bfaeSStefano Zampini xadj[1] = xadj_count; 4585c8587f34SStefano Zampini } else { 4586e7931f94SStefano Zampini if (xadj[1]) { 4587e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4588e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4589c8587f34SStefano Zampini } 4590e7931f94SStefano Zampini } 45913bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4592e7931f94SStefano Zampini if (use_square) { 4593e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4594e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4595e7931f94SStefano Zampini } 4596e7931f94SStefano Zampini } 4597e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4598e7931f94SStefano Zampini 45993837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4600e7931f94SStefano Zampini 4601e7931f94SStefano Zampini /* 4602e7931f94SStefano Zampini Restrict work on active processes only. 4603e7931f94SStefano Zampini */ 4604e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4605e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4606e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 46072b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4608d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4609e7931f94SStefano Zampini if (color) { 4610e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4611e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4612e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4613c8587f34SStefano Zampini } else { 461452e5ac9dSStefano Zampini Mat subdomain_adj; 461552e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 461652e5ac9dSStefano Zampini MatPartitioning partitioner; 461752e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 461852e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4619b0c7d250SStefano Zampini PetscBool aggregate; 4620b0c7d250SStefano Zampini 4621306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4622785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4623e7931f94SStefano Zampini prank = rank; 4624306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 46258002ef2cSStefano Zampini /* 4626e7931f94SStefano Zampini for (i=0;i<size;i++) { 4627e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4628c8587f34SStefano Zampini } 46298002ef2cSStefano Zampini */ 4630e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4631e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4632c8587f34SStefano Zampini } 4633e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4634b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4635b0c7d250SStefano Zampini if (aggregate) { 4636b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4637b0c7d250SStefano Zampini PetscMPIInt nrank; 4638b0c7d250SStefano Zampini PetscScalar *vals; 4639b0c7d250SStefano Zampini 4640b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4641b0c7d250SStefano Zampini lrows = 0; 4642b0c7d250SStefano Zampini if (nrank<redprocs) { 4643b0c7d250SStefano Zampini lrows = size/redprocs; 4644b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4645b0c7d250SStefano Zampini } 46465fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4647b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4648b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4649b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4650b0c7d250SStefano Zampini row = nrank; 4651b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4652b0c7d250SStefano Zampini cols = adjncy; 4653b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4654b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4655b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4656b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4657b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 465852e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 465952e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 466052e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4661b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4662b0c7d250SStefano Zampini } else { 4663306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4664b0c7d250SStefano Zampini } 466522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4666e7931f94SStefano Zampini 4667e7931f94SStefano Zampini /* Partition */ 4668306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4669e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4670e7931f94SStefano Zampini if (use_vwgt) { 46713837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4672e7931f94SStefano Zampini v_wgt[0] = local_size; 4673e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4674c8587f34SStefano Zampini } 467528143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 467628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4677e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4678e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 467922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4680e7931f94SStefano Zampini 468152e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 468252e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 468352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 468452e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4685b0c7d250SStefano Zampini if (!redprocs) { 4686b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 468728143c3dSStefano Zampini } else { 4688b0c7d250SStefano Zampini PetscInt idxs[1]; 4689b0c7d250SStefano Zampini PetscMPIInt tag; 4690b0c7d250SStefano Zampini MPI_Request *reqs; 4691b0c7d250SStefano Zampini 4692b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4693b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4694b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4695b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 469628143c3dSStefano Zampini } 4697b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4698b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4699b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4700b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4701e7931f94SStefano Zampini } 470252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4703e7931f94SStefano Zampini /* clean up */ 4704e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 470552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4706e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4707e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4708e7931f94SStefano Zampini } 4709e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4710e7931f94SStefano Zampini 4711e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4712e7931f94SStefano Zampini i = 1; 4713e7931f94SStefano Zampini if (color) i=0; 4714e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4715e7931f94SStefano Zampini /* get back IS */ 4716e7931f94SStefano Zampini *is_sends = ranks_send_to; 4717e7931f94SStefano Zampini PetscFunctionReturn(0); 4718e7931f94SStefano Zampini } 4719e7931f94SStefano Zampini 4720e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4721e7931f94SStefano Zampini 4722e7931f94SStefano Zampini #undef __FUNCT__ 4723e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 472453a05cb3SStefano 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[]) 4725e7931f94SStefano Zampini { 472670cf5478SStefano Zampini Mat local_mat; 4727e7931f94SStefano Zampini IS is_sends_internal; 47289d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 472928143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 47309d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4731e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4732e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4733e7931f94SStefano Zampini const PetscInt* is_indices; 4734e7931f94SStefano Zampini MatType new_local_type; 4735e7931f94SStefano Zampini /* buffers */ 4736e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 473728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 47389d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4739e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4740e7931f94SStefano Zampini /* MPI */ 474128143c3dSStefano Zampini MPI_Comm comm,comm_n; 474228143c3dSStefano Zampini PetscSubcomm subcomm; 4743e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 474428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 474528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 474628143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 474728143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 474828143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4749e7931f94SStefano Zampini PetscErrorCode ierr; 4750e7931f94SStefano Zampini 4751e7931f94SStefano Zampini PetscFunctionBegin; 475228143c3dSStefano Zampini /* TODO: add missing checks */ 475328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 475428143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 475528143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 475628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4757e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 475828143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4759e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4760e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4761e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4762e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4763e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 476428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 476570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 476670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 476728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 476870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 476970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 477070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 477170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 477270cf5478SStefano Zampini } 4773e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4774e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4775e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4776e7931f94SStefano Zampini if (!is_sends) { 477728143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4778b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4779c8587f34SStefano Zampini } else { 4780e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4781e7931f94SStefano Zampini is_sends_internal = is_sends; 4782c8587f34SStefano Zampini } 4783e7931f94SStefano Zampini 4784e7931f94SStefano Zampini /* get comm */ 4785a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4786e7931f94SStefano Zampini 4787e7931f94SStefano Zampini /* compute number of sends */ 4788e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4789e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4790e7931f94SStefano Zampini 4791e7931f94SStefano Zampini /* compute number of receives */ 4792e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4793785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4794e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4795e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4796e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4797e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4798e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4799e7931f94SStefano Zampini 480028143c3dSStefano Zampini /* restrict comm if requested */ 480128143c3dSStefano Zampini subcomm = 0; 480228143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 480328143c3dSStefano Zampini if (restrict_comm) { 4804779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4805779c1cceSStefano Zampini 480628143c3dSStefano Zampini color = 0; 480753a05cb3SStefano Zampini if (restrict_full) { 480853a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 480953a05cb3SStefano Zampini } else { 481053a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 481153a05cb3SStefano Zampini } 481228143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 481328143c3dSStefano Zampini subcommsize = commsize - subcommsize; 481428143c3dSStefano Zampini /* check if reuse has been requested */ 481528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 481628143c3dSStefano Zampini if (*mat_n) { 481728143c3dSStefano Zampini PetscMPIInt subcommsize2; 481828143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 481928143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 482028143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 482128143c3dSStefano Zampini } else { 482228143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 482328143c3dSStefano Zampini } 482428143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4825779c1cceSStefano Zampini PetscMPIInt rank; 4826779c1cceSStefano Zampini 4827779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 482828143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 482928143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 483028143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4831306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 483228143c3dSStefano Zampini } 483328143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 483428143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 483528143c3dSStefano Zampini } else { 483628143c3dSStefano Zampini comm_n = comm; 483728143c3dSStefano Zampini } 483828143c3dSStefano Zampini 4839e7931f94SStefano Zampini /* prepare send/receive buffers */ 4840785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4841e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4842785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4843e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 484428143c3dSStefano Zampini if (nis) { 4845854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 484628143c3dSStefano Zampini } 4847e7931f94SStefano Zampini 484828143c3dSStefano Zampini /* Get data from local matrices */ 4849e7931f94SStefano Zampini if (!isdense) { 4850a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4851e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4852e7931f94SStefano Zampini /* 4853e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4854e7931f94SStefano Zampini send_buffer_idxs should contain: 4855e7931f94SStefano Zampini - MatType_PRIVATE type 4856e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4857e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4858e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4859e7931f94SStefano Zampini */ 4860e7931f94SStefano Zampini } else { 4861e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 48623bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4863854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4864e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4865e7931f94SStefano Zampini send_buffer_idxs[1] = i; 48663bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4867e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 48683bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4869e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4870e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4871e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4872e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4873c8587f34SStefano Zampini } 4874c8587f34SStefano Zampini } 4875e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 487628143c3dSStefano Zampini /* additional is (if any) */ 487728143c3dSStefano Zampini if (nis) { 487828143c3dSStefano Zampini PetscMPIInt psum; 487928143c3dSStefano Zampini PetscInt j; 488028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 488128143c3dSStefano Zampini PetscInt plen; 488228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 488328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 488428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 488528143c3dSStefano Zampini } 4886854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 488728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 488828143c3dSStefano Zampini PetscInt plen; 488928143c3dSStefano Zampini const PetscInt *is_array_idxs; 489028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 489128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 489228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 489328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 489428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 489528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 489628143c3dSStefano Zampini } 489728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 489828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 489928143c3dSStefano Zampini } 490028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 490128143c3dSStefano Zampini } 490228143c3dSStefano Zampini 4903e7931f94SStefano Zampini buf_size_idxs = 0; 4904e7931f94SStefano Zampini buf_size_vals = 0; 490528143c3dSStefano Zampini buf_size_idxs_is = 0; 4906e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4907e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4908e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 490928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4910e7931f94SStefano Zampini } 4911785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4912785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 491395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4914e7931f94SStefano Zampini 4915e7931f94SStefano Zampini /* get new tags for clean communications */ 4916e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4917e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 491828143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4919e7931f94SStefano Zampini 4920e7931f94SStefano Zampini /* allocate for requests */ 4921785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4922785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 492395ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4924785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4925785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 492695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4927e7931f94SStefano Zampini 4928e7931f94SStefano Zampini /* communications */ 4929e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4930e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 493128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4932e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4933e7931f94SStefano Zampini source_dest = onodes[i]; 4934e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4935e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4936e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4937e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 493828143c3dSStefano Zampini if (nis) { 493928143c3dSStefano 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); 494028143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 494128143c3dSStefano Zampini } 4942e7931f94SStefano Zampini } 4943e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4944e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4945e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4946e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 494728143c3dSStefano Zampini if (nis) { 494828143c3dSStefano 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); 494928143c3dSStefano Zampini } 4950e7931f94SStefano Zampini } 4951e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4952e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4953e7931f94SStefano Zampini 4954e7931f94SStefano Zampini /* assemble new l2g map */ 4955e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4956e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49579d30be91SStefano Zampini new_local_rows = 0; 4958e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49599d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4960e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4961e7931f94SStefano Zampini } 49629d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4963e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49649d30be91SStefano Zampini new_local_rows = 0; 4965e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49669d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 49679d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4968e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4969e7931f94SStefano Zampini } 49709d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 49719d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4972e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4973e7931f94SStefano Zampini 4974e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4975e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4976e7931f94SStefano 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) */ 4977e7931f94SStefano Zampini if (n_recvs) { 497828143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4979e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4980e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4981e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4982e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4983e7931f94SStefano Zampini break; 4984e7931f94SStefano Zampini } 4985e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4986e7931f94SStefano Zampini } 4987e7931f94SStefano Zampini switch (new_local_type_private) { 498828143c3dSStefano Zampini case MATDENSE_PRIVATE: 498928143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4990e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4991e7931f94SStefano Zampini bs = 1; 499228143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 499328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 499428143c3dSStefano Zampini bs = 1; 499528143c3dSStefano Zampini } 4996e7931f94SStefano Zampini break; 4997e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4998e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4999e7931f94SStefano Zampini bs = 1; 5000e7931f94SStefano Zampini break; 5001e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5002e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5003e7931f94SStefano Zampini break; 5004e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5005e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5006e7931f94SStefano Zampini break; 5007e7931f94SStefano Zampini default: 50089d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5009e7931f94SStefano Zampini break; 5010e7931f94SStefano Zampini } 501128143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 501228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 501328143c3dSStefano Zampini bs = 1; 5014e7931f94SStefano Zampini } 5015e7931f94SStefano Zampini 501670cf5478SStefano Zampini /* create MATIS object if needed */ 501770cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5018e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5019e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 502070cf5478SStefano Zampini } else { 502170cf5478SStefano Zampini /* it also destroys the local matrices */ 502270cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 502370cf5478SStefano Zampini } 502470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5025e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 50269d30be91SStefano Zampini 50279d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 50289d30be91SStefano Zampini 50299d30be91SStefano Zampini /* Global to local map of received indices */ 50309d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 50319d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 50329d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 50339d30be91SStefano Zampini 50349d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 50359d30be91SStefano Zampini buf_size_idxs = 0; 50369d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50379d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 50389d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 50399d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 50409d30be91SStefano Zampini } 50419d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 50429d30be91SStefano Zampini 50439d30be91SStefano Zampini /* set preallocation */ 50449d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 50459d30be91SStefano Zampini if (!newisdense) { 50469d30be91SStefano Zampini PetscInt *new_local_nnz=0; 50479d30be91SStefano Zampini 50489d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 50499d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 50509d30be91SStefano Zampini if (n_recvs) { 50519d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 50529d30be91SStefano Zampini } 50539d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50549d30be91SStefano Zampini PetscInt j; 50559d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 50569d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 50579d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 50589d30be91SStefano Zampini } 50599d30be91SStefano Zampini } else { 50609d30be91SStefano Zampini /* TODO */ 50619d30be91SStefano Zampini } 50629d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 50639d30be91SStefano Zampini } 50649d30be91SStefano Zampini if (new_local_nnz) { 50659d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 50669d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 50679d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 50689d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50699d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 50709d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50719d30be91SStefano Zampini } else { 50729d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50739d30be91SStefano Zampini } 50749d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 50759d30be91SStefano Zampini } else { 50769d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50779d30be91SStefano Zampini } 5078e7931f94SStefano Zampini 5079e7931f94SStefano Zampini /* set values */ 5080e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 50819d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5082e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5083e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5084e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 50859d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5086e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5087e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5088e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 508928143c3dSStefano Zampini } else { 509028143c3dSStefano Zampini /* TODO */ 5091e7931f94SStefano Zampini } 5092e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5093e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5094e7931f94SStefano Zampini } 5095e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5096e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 509770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 509870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 50999d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 51009d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5101e7931f94SStefano Zampini 5102dfd14d43SStefano Zampini #if 0 510328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5104e7931f94SStefano Zampini Vec lvec,rvec; 5105e7931f94SStefano Zampini PetscReal infty_error; 5106e7931f94SStefano Zampini 51072a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5108e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5109e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5110e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 511170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5112e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5113e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5114e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5115e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5116e7931f94SStefano Zampini } 511728143c3dSStefano Zampini #endif 5118e7931f94SStefano Zampini 511928143c3dSStefano Zampini /* assemble new additional is (if any) */ 512028143c3dSStefano Zampini if (nis) { 512128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 512228143c3dSStefano Zampini 512328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5124854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 512528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 512628143c3dSStefano Zampini psum = 0; 512728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 512828143c3dSStefano Zampini for (j=0;j<nis;j++) { 512928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 513028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 513128143c3dSStefano Zampini psum += plen; 513228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 513328143c3dSStefano Zampini } 513428143c3dSStefano Zampini } 5135854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5136854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 513728143c3dSStefano Zampini for (i=1;i<nis;i++) { 513828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 513928143c3dSStefano Zampini } 514028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 514128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 514228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 514328143c3dSStefano Zampini for (j=0;j<nis;j++) { 514428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 514528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 514628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 514728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 514828143c3dSStefano Zampini } 514928143c3dSStefano Zampini } 515028143c3dSStefano Zampini for (i=0;i<nis;i++) { 515128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 515228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 515328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 515428143c3dSStefano Zampini } 515528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 515628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 515728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 515828143c3dSStefano Zampini } 5159e7931f94SStefano Zampini /* free workspace */ 516028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5161e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5162e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5163e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5164e7931f94SStefano Zampini if (isdense) { 5165e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5166e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5167e7931f94SStefano Zampini } else { 5168e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5169e7931f94SStefano Zampini } 517028143c3dSStefano Zampini if (nis) { 517128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 517228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 517328143c3dSStefano Zampini } 5174e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5175e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 517628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5177e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5178e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 517928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5180e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5181e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5182e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5183e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5184e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 518528143c3dSStefano Zampini if (nis) { 518628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 518728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 518828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 518928143c3dSStefano Zampini } 519028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 519128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 519228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 519328143c3dSStefano Zampini for (i=0;i<nis;i++) { 519428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 519528143c3dSStefano Zampini } 519653a05cb3SStefano Zampini *mat_n = NULL; 519728143c3dSStefano Zampini } 5198e7931f94SStefano Zampini PetscFunctionReturn(0); 5199e7931f94SStefano Zampini } 5200a57a6d2fSStefano Zampini 520112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5202af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 520312edc857SStefano Zampini 5204c8587f34SStefano Zampini #undef __FUNCT__ 5205c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5206c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5207c8587f34SStefano Zampini { 5208c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5209c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 521020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 52119881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 521220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 52136e683305SStefano Zampini IS coarse_is,*isarray; 52146e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 521530368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5216f9eb5b7dSStefano Zampini PC pc_temp; 5217c8587f34SStefano Zampini PCType coarse_pc_type; 5218c8587f34SStefano Zampini KSPType coarse_ksp_type; 5219f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 52204f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 52216e683305SStefano Zampini Mat t_coarse_mat_is; 52226e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 52236e683305SStefano Zampini PetscMPIInt all_procs; 522474e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 522568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 522622bc73bbSStefano Zampini PetscScalar *array; 52279881197aSStefano Zampini PetscErrorCode ierr; 5228fdc09c96SStefano Zampini 5229c8587f34SStefano Zampini PetscFunctionBegin; 5230c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 523168457ee5SStefano 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 */ 5232fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 52335a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5234fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5235f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5236f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5237f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5238fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 523951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 524051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5241dc4bcba2SStefano Zampini PC pc; 5242dc4bcba2SStefano Zampini PetscBool isbddc; 5243dc4bcba2SStefano Zampini 5244dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5245dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5246dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5247dc4bcba2SStefano Zampini if (isbddc) { 524863c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 524963c961adSStefano Zampini } else { 5250727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 525163c961adSStefano Zampini } 5252fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5253fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5254fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5255f4ddd8eeSStefano Zampini } 5256fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5257fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5258f4ddd8eeSStefano Zampini } 525970cf5478SStefano Zampini /* reset any subassembling information */ 526070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 52616e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 52626e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5263fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5264f4ddd8eeSStefano Zampini } 5265c8587f34SStefano Zampini 52666e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 52672b510759SStefano Zampini im_active = !!(pcis->n); 52682b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 52696e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 52706e683305SStefano Zampini void_procs = all_procs-active_procs; 52716e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 527274e2c79eSStefano Zampini redist = PETSC_FALSE; 527322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 52746e683305SStefano Zampini csin_ml = PETSC_TRUE; 52756e683305SStefano Zampini ncoarse_ml = void_procs; 5276779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5277779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 52786e683305SStefano Zampini csin_ds = PETSC_TRUE; 527918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 528018a45a71SStefano Zampini redist = PETSC_TRUE; 528118a45a71SStefano Zampini } else { 52826e683305SStefano Zampini csin_ds = PETSC_TRUE; 5283779c1cceSStefano Zampini ncoarse_ds = active_procs; 5284779c1cceSStefano Zampini redist = PETSC_TRUE; 528518a45a71SStefano Zampini } 52866e683305SStefano Zampini } else { 52876e683305SStefano Zampini csin_ml = PETSC_FALSE; 52886e683305SStefano Zampini ncoarse_ml = all_procs; 52896e683305SStefano Zampini if (void_procs) { 52906e683305SStefano Zampini csin_ds = PETSC_TRUE; 52916e683305SStefano Zampini ncoarse_ds = void_procs; 52926e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 52936e683305SStefano Zampini } else { 5294779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 529574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 529674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 529774e2c79eSStefano Zampini redist = PETSC_TRUE; 529874e2c79eSStefano Zampini } else { 52996e683305SStefano Zampini csin_ds = PETSC_FALSE; 53006e683305SStefano Zampini ncoarse_ds = all_procs; 53016e683305SStefano Zampini } 53026e683305SStefano Zampini } 530374e2c79eSStefano Zampini } 53046e683305SStefano Zampini 53056e683305SStefano Zampini /* 53066e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 53076e683305SStefano Zampini - we have not exceeded the number of levels requested 53086e683305SStefano Zampini - we can actually subassemble the active processes 53096e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 53106e683305SStefano Zampini */ 53116e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 53126e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 53136e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 53146e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 53156e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5316f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 53172b510759SStefano Zampini } else { 5318f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5319c8587f34SStefano Zampini } 5320c8587f34SStefano Zampini } 53216e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 53226e683305SStefano Zampini if (multilevel_allowed) { 53236e683305SStefano Zampini ncoarse = ncoarse_ml; 53246e683305SStefano Zampini csin = csin_ml; 532558da7f69SStefano Zampini redist = PETSC_FALSE; 53266e683305SStefano Zampini } else { 53276e683305SStefano Zampini ncoarse = ncoarse_ds; 53286e683305SStefano Zampini csin = csin_ds; 53296e683305SStefano Zampini } 5330e7931f94SStefano Zampini 5331abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5332abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5333abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5334abbbba34SStefano Zampini 5335abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 533622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 533722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 533822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 533922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5340e176bc59SStefano 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); 53416e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 53426e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53436e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5344abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5345abbbba34SStefano Zampini 53466e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 534730368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 53486e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 53496e683305SStefano Zampini const PetscInt *idxs; 53506e683305SStefano Zampini ISLocalToGlobalMapping tmap; 53516e683305SStefano Zampini 53526e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 53530be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 53546e683305SStefano Zampini /* allocate space for temporary storage */ 5355854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5356854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 53576e683305SStefano Zampini /* allocate for IS array */ 53586e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 53596e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 536030368db7SStefano Zampini nisvert = 0; 536130368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 536230368db7SStefano Zampini nisvert = 1; 536330368db7SStefano Zampini } 536430368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5365854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 53666e683305SStefano Zampini /* dofs splitting */ 53676e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 53686e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 53696e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 53706e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53716e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53726e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53736e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 537430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 53756e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 53766e683305SStefano Zampini } 53776e683305SStefano Zampini /* neumann boundaries */ 53786e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 53796e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 53806e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 53816e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53826e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53836e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53846e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 538530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 53866e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 53876e683305SStefano Zampini } 538830368db7SStefano Zampini /* primal vertices (benign) */ 538930368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 539030368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 539130368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 539230368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 539330368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 539430368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 539530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 539630368db7SStefano Zampini } 53976e683305SStefano Zampini /* free memory */ 53986e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 53996e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 54006e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 54016e683305SStefano Zampini } else { 54026e683305SStefano Zampini nis = 0; 54036e683305SStefano Zampini nisdofs = 0; 54046e683305SStefano Zampini nisneu = 0; 540530368db7SStefano Zampini nisvert = 0; 54066e683305SStefano Zampini isarray = NULL; 54076e683305SStefano Zampini } 54086e683305SStefano Zampini /* destroy no longer needed map */ 54096e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 54106e683305SStefano Zampini 54116e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 54126e683305SStefano Zampini coarse_mat_is = NULL; 54136e683305SStefano Zampini if (csin) { 54146e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 541574e2c79eSStefano Zampini if (redist) { 541674e2c79eSStefano Zampini PetscMPIInt rank; 5417779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 541874e2c79eSStefano Zampini 541974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 542058da7f69SStefano Zampini spc = active_procs/ncoarse; 542158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5422779c1cceSStefano Zampini if (im_active) { 5423779c1cceSStefano Zampini destsize = 1; 542474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 542574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 542674e2c79eSStefano Zampini } else { 542774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 542874e2c79eSStefano Zampini } 542974e2c79eSStefano Zampini } else { 5430779c1cceSStefano Zampini destsize = 0; 54316e683305SStefano Zampini } 5432779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5433779c1cceSStefano Zampini } else if (csin_type_simple) { 54346e683305SStefano Zampini PetscMPIInt rank; 54356e683305SStefano Zampini PetscInt issize,isidx; 5436779c1cceSStefano Zampini 54376e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 54386e683305SStefano Zampini if (im_active) { 54396e683305SStefano Zampini issize = 1; 54406e683305SStefano Zampini isidx = (PetscInt)rank; 54416e683305SStefano Zampini } else { 54426e683305SStefano Zampini issize = 0; 54436e683305SStefano Zampini isidx = -1; 54446e683305SStefano Zampini } 54456e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5446779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5447b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54486e683305SStefano Zampini } 5449779c1cceSStefano Zampini 5450779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5451779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5452779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5453779c1cceSStefano Zampini PetscInt *coarse_candidates; 5454779c1cceSStefano Zampini const PetscInt* tisindices; 5455779c1cceSStefano Zampini 5456779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5457779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5458779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5459779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5460779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5461779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5462779c1cceSStefano Zampini } 5463779c1cceSStefano Zampini } 5464779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5465779c1cceSStefano Zampini 5466779c1cceSStefano Zampini 54676e683305SStefano Zampini if (pcbddc->dbg_flag) { 54686e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54696e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 54706e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54716e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5472779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 54736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 54746e683305SStefano Zampini } 54756e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 54766e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54776e683305SStefano Zampini } 54786e683305SStefano Zampini /* shift the pattern on coarse candidates */ 54796e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 54806e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5481854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 54826e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 54836e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 54846e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 54856e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 54866e683305SStefano Zampini } 54876e683305SStefano Zampini if (pcbddc->dbg_flag) { 54886e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54896e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 54906e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54916e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54926e683305SStefano Zampini } 5493779c1cceSStefano Zampini } 54946e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 549553a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 549653a05cb3SStefano 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); 549753a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 549853a05cb3SStefano 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); 549953a05cb3SStefano Zampini } 55006e683305SStefano Zampini } else { 55016e683305SStefano Zampini if (pcbddc->dbg_flag) { 55026e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 55046e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55056e683305SStefano Zampini } 55066e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 55076e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 55086e683305SStefano Zampini } 55096e683305SStefano Zampini 55106e683305SStefano Zampini /* create local to global scatters for coarse problem */ 551168457ee5SStefano Zampini if (compute_vecs) { 55126e683305SStefano Zampini PetscInt lrows; 55136e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 55146e683305SStefano Zampini if (coarse_mat_is) { 55156e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 55166e683305SStefano Zampini } else { 55176e683305SStefano Zampini lrows = 0; 55186e683305SStefano Zampini } 55196e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 55206e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 55216e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 55226e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55236e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55246e683305SStefano Zampini } 55256e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 55266e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5527c8587f34SStefano Zampini 5528f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5529f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5530f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5531f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5532f9eb5b7dSStefano Zampini } else { 5533f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5534f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5535c8587f34SStefano Zampini } 5536c8587f34SStefano Zampini 55376e683305SStefano Zampini /* print some info if requested */ 55386e683305SStefano Zampini if (pcbddc->dbg_flag) { 55396e683305SStefano Zampini if (!multilevel_allowed) { 55406e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55416e683305SStefano Zampini if (multilevel_requested) { 55426e683305SStefano 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); 55436e683305SStefano Zampini } else if (pcbddc->max_levels) { 55446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 55456e683305SStefano Zampini } 55466e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55476e683305SStefano Zampini } 55486e683305SStefano Zampini } 55496e683305SStefano Zampini 5550f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 55516e683305SStefano Zampini if (coarse_mat_is) { 55526e683305SStefano Zampini MatReuse coarse_mat_reuse; 55536a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 55546e683305SStefano Zampini if (pcbddc->dbg_flag) { 55556e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 55566e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 55576e683305SStefano Zampini } 5558f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5559312be037SStefano Zampini char prefix[256],str_level[16]; 5560e604994aSStefano Zampini size_t len; 55616e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5562422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5563c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5564f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 55655f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5566c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 55676e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5568c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5569c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5570e604994aSStefano Zampini /* prefix */ 5571e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5572e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5573e604994aSStefano Zampini if (!pcbddc->current_level) { 5574e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5575e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5576c8587f34SStefano Zampini } else { 5577e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5578312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5579312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 558034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5581312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5582e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5583e604994aSStefano Zampini } 5584e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 55853e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 55863e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 55873e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 55883e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5589f9eb5b7dSStefano Zampini /* allow user customization */ 5590f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 55913e3c6dadSStefano Zampini } 55923e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 559351bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 55943e3c6dadSStefano Zampini if (nisdofs) { 55953e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 55963e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 55973e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 55983e3c6dadSStefano Zampini } 55993e3c6dadSStefano Zampini } 56003e3c6dadSStefano Zampini if (nisneu) { 56013e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 56023e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5603312be037SStefano Zampini } 560430368db7SStefano Zampini if (nisvert) { 560530368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 560630368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 560730368db7SStefano Zampini } 5608f9eb5b7dSStefano Zampini 5609f9eb5b7dSStefano Zampini /* get some info after set from options */ 5610f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5611f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 56124f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 56136e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5614f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5615f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5616f9eb5b7dSStefano Zampini } 561739f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 56184f3a063dSStefano Zampini if (isredundant) { 56194f3a063dSStefano Zampini KSP inner_ksp; 56204f3a063dSStefano Zampini PC inner_pc; 56214f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 56224f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 56234f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 56244f3a063dSStefano Zampini } 5625f9eb5b7dSStefano Zampini 5626f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5627fa7f1dd8SStefano Zampini if (coarse_reuse) { 562881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5629fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 56306e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5631fa7f1dd8SStefano Zampini } else { 56326e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5633fa7f1dd8SStefano Zampini } 5634c8587f34SStefano Zampini if (isbddc || isnn) { 5635d4d8cf7bSStefano Zampini if (isbddc) { /* currently there's no API for this */ 5636d4d8cf7bSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc_temp->data; 5637d4d8cf7bSStefano Zampini pcbddc->detect_disconnected = PETSC_TRUE; 5638d4d8cf7bSStefano Zampini } 563922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 564070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5641b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 564222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 56436e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56446e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 56456e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 56466e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 564722b6e8a2SStefano Zampini } 564870cf5478SStefano Zampini } 564953a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 565070cf5478SStefano Zampini } else { 565122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 565222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 565322bc73bbSStefano Zampini } 565422bc73bbSStefano Zampini } else { 56552e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5656c8587f34SStefano Zampini } 5657c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5658c8587f34SStefano Zampini 56593301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 56605a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 56613301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 56623301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 56633301b35fSStefano Zampini } 56643301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 56653301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 56663301b35fSStefano Zampini } 56673301b35fSStefano Zampini if (pc->pmat->spd_set) { 56683301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 56693301b35fSStefano Zampini } 56706e683305SStefano Zampini /* set operators */ 56715f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 56726e683305SStefano Zampini if (pcbddc->dbg_flag) { 56736e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56746e683305SStefano Zampini } 56756e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 56766e683305SStefano Zampini coarse_mat = 0; 56776e683305SStefano Zampini } 56786e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5679b1ecc7b1SStefano Zampini #if 0 5680b9b85e73SStefano Zampini { 5681b9b85e73SStefano Zampini PetscViewer viewer; 5682b9b85e73SStefano Zampini char filename[256]; 5683b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5684b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5685b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5686b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5687b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5688b9b85e73SStefano Zampini } 5689b9b85e73SStefano Zampini #endif 5690c8587f34SStefano Zampini 5691c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5692298c0119SStefano Zampini #if 0 5693c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5694c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 569598a51de6SStefano Zampini } 5696298c0119SStefano Zampini #endif 5697b0f5fe93SStefano Zampini /* hack */ 569898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 569998a51de6SStefano Zampini Vec crhs,csol; 570004708bb6SStefano Zampini 5701f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5702f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5703f347579bSStefano Zampini if (!csol) { 57042a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5705f9eb5b7dSStefano Zampini } 5706f347579bSStefano Zampini if (!crhs) { 57072a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5708f347579bSStefano Zampini } 5709b0f5fe93SStefano Zampini } 5710b0f5fe93SStefano Zampini 5711b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5712b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5713b0f5fe93SStefano Zampini 5714b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 57154f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 57164f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 57174f1b2e48SStefano Zampini } 5718b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5719b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5720b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5721b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5722b0f5fe93SStefano Zampini if (coarse_mat) { 5723b0f5fe93SStefano Zampini Vec nullv; 5724b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5725b0f5fe93SStefano Zampini PetscInt nl; 5726b0f5fe93SStefano Zampini 5727b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5728b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5729b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5730b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5731b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5732b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5733b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5734b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5735b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5736b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5737b0f5fe93SStefano Zampini } 5738b0f5fe93SStefano Zampini } 5739b0f5fe93SStefano Zampini 5740b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5741b0f5fe93SStefano Zampini PetscBool ispreonly; 5742b0f5fe93SStefano Zampini 5743b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5744b0f5fe93SStefano Zampini PetscBool isnull; 5745b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5746d4d8cf7bSStefano Zampini if (0) { 574730368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5748b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5749b0f5fe93SStefano Zampini } else { 5750b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5751b0f5fe93SStefano Zampini } 5752b0f5fe93SStefano Zampini } else { 5753b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5754b0f5fe93SStefano Zampini } 5755b0f5fe93SStefano Zampini } 5756b0f5fe93SStefano Zampini /* setup coarse ksp */ 5757b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5758cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5759cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 57606e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5761c8587f34SStefano Zampini KSP check_ksp; 57622b510759SStefano Zampini KSPType check_ksp_type; 5763c8587f34SStefano Zampini PC check_pc; 57646e683305SStefano Zampini Vec check_vec,coarse_vec; 57656a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 57662b510759SStefano Zampini PetscInt its; 57676e683305SStefano Zampini PetscBool compute_eigs; 57686e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 57696e683305SStefano Zampini PetscInt neigs; 57708e185a42SStefano Zampini const char *prefix; 5771c8587f34SStefano Zampini 57722b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 57736e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5774422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 577523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5776f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 57772b510759SStefano Zampini if (ispreonly) { 57782b510759SStefano Zampini check_ksp_type = KSPPREONLY; 57796e683305SStefano Zampini compute_eigs = PETSC_FALSE; 57802b510759SStefano Zampini } else { 5781cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 57826e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5783c8587f34SStefano Zampini } 5784c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 57856e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 57866e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 57876e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5788a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5789a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5790a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5791a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5792c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5793c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5794c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5795c8587f34SStefano Zampini /* create random vec */ 57966e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 57976e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5798c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5799c8587f34SStefano Zampini if (CoarseNullSpace) { 5800c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5801c8587f34SStefano Zampini } 58026e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5803c8587f34SStefano Zampini /* solve coarse problem */ 58046e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5805c8587f34SStefano Zampini if (CoarseNullSpace) { 58066e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5807c8587f34SStefano Zampini } 5808cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 58096e683305SStefano Zampini if (compute_eigs) { 5810854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5811854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 58126e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 58136e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 58146e683305SStefano Zampini lambda_min = eigs_r[0]; 58156e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 58166e683305SStefano Zampini if (lambda_max>lambda_min) { 5817cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5818cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5819cbcc2c2aSStefano Zampini } 5820c8587f34SStefano Zampini } 5821c8587f34SStefano Zampini } 5822cbcc2c2aSStefano Zampini 5823c8587f34SStefano Zampini /* check coarse problem residual error */ 58246e683305SStefano Zampini if (pcbddc->dbg_flag) { 58256e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 58266e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58276e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5828c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 58296e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 58306e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5831c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5832779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 58336e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 58346e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 58356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 58366e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5837b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5838b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5839b0f5fe93SStefano Zampini } 58406e683305SStefano Zampini if (compute_eigs) { 58416e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5842deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5843c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 58446e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 58456e683305SStefano 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); 58466e683305SStefano Zampini for (i=0;i<neigs;i++) { 58476e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5848c8587f34SStefano Zampini } 58496e683305SStefano Zampini } 58506e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 58516e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58526e683305SStefano Zampini } 5853c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 58546e683305SStefano Zampini if (compute_eigs) { 58556e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 58566e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5857c8587f34SStefano Zampini } 58586e683305SStefano Zampini } 58596e683305SStefano Zampini } 5860cbcc2c2aSStefano Zampini /* print additional info */ 5861cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 58626e683305SStefano Zampini /* waits until all processes reaches this point */ 58636e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5864cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5865cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5866cbcc2c2aSStefano Zampini } 5867cbcc2c2aSStefano Zampini 58682b510759SStefano Zampini /* free memory */ 5869c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5870fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5871c8587f34SStefano Zampini PetscFunctionReturn(0); 5872c8587f34SStefano Zampini } 5873674ae819SStefano Zampini 5874f34684f1SStefano Zampini #undef __FUNCT__ 5875f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5876f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5877f34684f1SStefano Zampini { 5878f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5879f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5880f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5881dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5882dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 588373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5884dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5885f34684f1SStefano Zampini PetscErrorCode ierr; 5886f34684f1SStefano Zampini 5887f34684f1SStefano Zampini PetscFunctionBegin; 5888f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 58890e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 58900e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5891727cdba6SStefano Zampini } 5892dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 58933bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5894dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5895dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5896dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5897dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5898dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5899dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 59000e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 59010e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 59020e6343abSStefano Zampini } 5903dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5904dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5905dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5906dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5907dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5908f34684f1SStefano Zampini 5909f34684f1SStefano Zampini /* check numbering */ 5910f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5911019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5912dc456d91SStefano Zampini PetscInt i; 5913b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5914f34684f1SStefano Zampini 5915f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5916f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5917f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 59181575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5919019a44ceSStefano Zampini /* counter */ 5920019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5921019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5922019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5923019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5924019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5925019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5926f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5927f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5928727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5929f34684f1SStefano Zampini } 5930f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5931f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5932f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5933e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5934e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5935e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5936e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5937f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5938019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5939f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5940019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 594175c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 594275c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5943b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5944019a44ceSStefano 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); 5945f34684f1SStefano Zampini } 5946f34684f1SStefano Zampini } 5947019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5948b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5949f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5950f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5951f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5952f34684f1SStefano Zampini } 5953f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5954f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5955e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5956e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5957f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5958f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5959b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5960ca8b9ea9SStefano Zampini PetscInt *gidxs; 5961ca8b9ea9SStefano Zampini 5962ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 59633bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5964f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5965f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5966f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5967f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 59684bc2dc4bSStefano 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); 5969f34684f1SStefano Zampini } 5970f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5971ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5972f34684f1SStefano Zampini } 5973f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59741575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5975302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5976f34684f1SStefano Zampini } 59778bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5978f34684f1SStefano Zampini /* get back data */ 5979f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5980f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5981674ae819SStefano Zampini PetscFunctionReturn(0); 5982674ae819SStefano Zampini } 5983674ae819SStefano Zampini 5984e456f2a8SStefano Zampini #undef __FUNCT__ 5985e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5986a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5987e456f2a8SStefano Zampini { 5988e456f2a8SStefano Zampini IS localis_t; 5989a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5990e456f2a8SStefano Zampini PetscScalar *vals; 5991e456f2a8SStefano Zampini PetscErrorCode ierr; 5992e456f2a8SStefano Zampini 5993e456f2a8SStefano Zampini PetscFunctionBegin; 5994a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5995e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5996854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5997e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5998e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5999a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6000a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 60011035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6002a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 60031035eff8SStefano Zampini } 6004a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6005e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6006e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6007a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6008a7dc3881SStefano Zampini /* now compute set in local ordering */ 6009a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6010a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6011a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6012a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6013a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6014ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6015e456f2a8SStefano Zampini lsize++; 6016e456f2a8SStefano Zampini } 6017e456f2a8SStefano Zampini } 6018854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6019a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6020ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6021e456f2a8SStefano Zampini idxs[lsize++] = i; 6022e456f2a8SStefano Zampini } 6023e456f2a8SStefano Zampini } 6024a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6025a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6026e456f2a8SStefano Zampini *localis = localis_t; 6027e456f2a8SStefano Zampini PetscFunctionReturn(0); 6028e456f2a8SStefano Zampini } 6029906d46d4SStefano Zampini 6030906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6031906d46d4SStefano Zampini #undef __FUNCT__ 6032906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6033906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6034906d46d4SStefano Zampini { 6035906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6036906d46d4SStefano Zampini PetscErrorCode ierr; 6037906d46d4SStefano Zampini 6038906d46d4SStefano Zampini PetscFunctionBegin; 6039906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6040906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6041906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6042906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6043906d46d4SStefano Zampini PetscFunctionReturn(0); 6044906d46d4SStefano Zampini } 6045906d46d4SStefano Zampini 6046906d46d4SStefano Zampini #undef __FUNCT__ 6047906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6048906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6049906d46d4SStefano Zampini { 6050906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6051906d46d4SStefano Zampini PetscErrorCode ierr; 6052906d46d4SStefano Zampini 6053906d46d4SStefano Zampini PetscFunctionBegin; 6054906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6055906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6056906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6057906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6058906d46d4SStefano Zampini PetscFunctionReturn(0); 6059906d46d4SStefano Zampini } 6060b96c3477SStefano Zampini 6061b96c3477SStefano Zampini #undef __FUNCT__ 6062b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 606308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6064b96c3477SStefano Zampini { 6065a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6066b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6067b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6068a64f4aa4SStefano Zampini Mat S_j; 6069b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6070b96c3477SStefano Zampini PetscBool free_used_adj; 6071b96c3477SStefano Zampini PetscErrorCode ierr; 6072b96c3477SStefano Zampini 6073b96c3477SStefano Zampini PetscFunctionBegin; 6074b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6075b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 607608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6077b96c3477SStefano Zampini used_xadj = NULL; 6078b96c3477SStefano Zampini used_adjncy = NULL; 6079b96c3477SStefano Zampini } else { 608008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 608108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 608208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 608308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6084b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6085b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6086b96c3477SStefano Zampini } else { 60872fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6088b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6089b96c3477SStefano Zampini PetscInt nvtxs; 6090b96c3477SStefano Zampini 60912fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 60922fffb893SStefano Zampini if (flg_row) { 6093b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6094b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6095b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6096b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 60972fffb893SStefano Zampini } else { 60982fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 60992fffb893SStefano Zampini used_xadj = NULL; 61002fffb893SStefano Zampini used_adjncy = NULL; 61012fffb893SStefano Zampini } 61022fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6103b96c3477SStefano Zampini } 6104b96c3477SStefano Zampini } 6105d5574798SStefano Zampini 6106d5574798SStefano Zampini /* setup sub_schurs data */ 6107a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6108df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6109df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6110a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6111ca92afb2SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL);CHKERRQ(ierr); 6112a64f4aa4SStefano Zampini } else { 61136816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 611404708bb6SStefano Zampini PetscBool isseqaij; 6115a3df083aSStefano Zampini PetscInt benign_n; 6116a3df083aSStefano Zampini 61175feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 61185feab87aSStefano Zampini PetscInt n_vertices; 61195feab87aSStefano Zampini 61205feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 61212034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 61225feab87aSStefano Zampini } 612304708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 612404708bb6SStefano Zampini if (!isseqaij) { 612504708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 612604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 612704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 612804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 612904708bb6SStefano Zampini } else { 613004708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 613104708bb6SStefano Zampini } 613204708bb6SStefano Zampini } 6133a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6134a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6135ca92afb2SStefano Zampini } else { 6136a3df083aSStefano Zampini benign_n = 0; 6137ca92afb2SStefano Zampini } 6138a3df083aSStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 613988c03ad3SStefano Zampini /* if the velocities are in the original basis, then we need to eliminate the no-net-flux from the S_Ej 614088c03ad3SStefano Zampini we assume that the user has passed in the associated quadrature weights in the near-null space attached to pc->pmat 614188c03ad3SStefano Zampini We need to compute the change of basis according to the quadrature weights, and this could not be done, at the moment, without some hacking */ 6142*fc227af8SStefano Zampini if (pcbddc->benign_saddle_point && !pcbddc->use_change_of_basis && !pcbddc->user_ChangeOfBasisMatrix) { 614388c03ad3SStefano Zampini PC_IS *pcisf; 614488c03ad3SStefano Zampini PC_BDDC *pcbddcf; 614588c03ad3SStefano Zampini PC pcf; 614688c03ad3SStefano Zampini 614788c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 614888c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 614988c03ad3SStefano Zampini } 615088c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 615188c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 615288c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 615388c03ad3SStefano Zampini /* hacks */ 615488c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 615588c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 615688c03ad3SStefano Zampini *pcisf = *pcis; 615788c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 615888c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 615988c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 616088c03ad3SStefano Zampini pcbddcf->sub_schurs = sub_schurs; 616188c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 616288c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 616388c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 616488c03ad3SStefano Zampini pcbddcf->use_deluxe_scaling = PETSC_TRUE; 616588c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 616688c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 616788c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 616888c03ad3SStefano Zampini pcbddcf->sub_schurs = NULL; 616988c03ad3SStefano Zampini ierr = MatDestroy(&pcbddcf->ConstraintMatrix);CHKERRQ(ierr); 617088c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 617188c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 617288c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 617388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 617488c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 617588c03ad3SStefano Zampini pcf->ops->destroy = NULL; 617688c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 617788c03ad3SStefano Zampini } 6178ca92afb2SStefano Zampini } 6179d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6180b96c3477SStefano Zampini 6181b96c3477SStefano Zampini /* free adjacency */ 6182b96c3477SStefano Zampini if (free_used_adj) { 6183b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6184b96c3477SStefano Zampini } 6185b96c3477SStefano Zampini PetscFunctionReturn(0); 6186b96c3477SStefano Zampini } 6187b96c3477SStefano Zampini 6188b96c3477SStefano Zampini #undef __FUNCT__ 6189b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 619008122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6191b96c3477SStefano Zampini { 6192b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6193b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6194b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6195b96c3477SStefano Zampini PCBDDCGraph graph; 6196b96c3477SStefano Zampini PetscErrorCode ierr; 6197b96c3477SStefano Zampini 6198b96c3477SStefano Zampini PetscFunctionBegin; 6199b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 620008122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 62013301b35fSStefano Zampini IS verticesIS,verticescomm; 62023301b35fSStefano Zampini PetscInt vsize,*idxs; 6203b96c3477SStefano Zampini 6204b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 62053301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 62063301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62073301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 62083301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62093301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6210b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 62117fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 62123301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 62133301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6214b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6215b96c3477SStefano Zampini /* 6216b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6217b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6218b96c3477SStefano Zampini } 6219b96c3477SStefano Zampini */ 6220b96c3477SStefano Zampini } else { 6221b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6222b96c3477SStefano Zampini } 6223b96c3477SStefano Zampini 6224b96c3477SStefano Zampini /* sub_schurs init */ 6225a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6226a64f4aa4SStefano Zampini 6227b96c3477SStefano Zampini /* free graph struct */ 622808122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6229b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6230b96c3477SStefano Zampini } 6231b96c3477SStefano Zampini PetscFunctionReturn(0); 6232b96c3477SStefano Zampini } 6233fa34dd3eSStefano Zampini 6234fa34dd3eSStefano Zampini #undef __FUNCT__ 6235fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6236fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6237fa34dd3eSStefano Zampini { 6238fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6239fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6240fa34dd3eSStefano Zampini PetscErrorCode ierr; 6241fa34dd3eSStefano Zampini 6242fa34dd3eSStefano Zampini PetscFunctionBegin; 6243fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6244fa34dd3eSStefano Zampini IS zerodiag = NULL; 62454f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6246fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 62474f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 624875c01103SStefano Zampini PetscReal norm; 6249fa34dd3eSStefano Zampini PetscInt i; 6250fa34dd3eSStefano Zampini 6251fa34dd3eSStefano Zampini /* B0 and B0_B */ 6252fa34dd3eSStefano Zampini if (zerodiag) { 6253fa34dd3eSStefano Zampini IS dummy; 6254fa34dd3eSStefano Zampini 62554f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 62564f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6257fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6258fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6259fa34dd3eSStefano Zampini } 6260fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6261fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6262fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6263fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6264fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6265fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6266fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6267fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6268fa34dd3eSStefano Zampini /* S_j */ 6269fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6270fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6271fa34dd3eSStefano Zampini 6272fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6273fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6274fa34dd3eSStefano Zampini /* continuous in primal space */ 6275fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6276fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6277fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6278fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 62794f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 62804f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6281fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6282fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6283fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6284fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6285fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6286fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6287fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6288fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6289fa34dd3eSStefano Zampini 6290fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6291fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6292fa34dd3eSStefano Zampini /* local with Schur */ 6293fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6294fa34dd3eSStefano Zampini if (zerodiag) { 6295fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62964f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6297fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6298fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6299fa34dd3eSStefano Zampini } 6300fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6301fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6302fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6303fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6304fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6305fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6306fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6307fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6308fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6309fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6310fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6311fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6312fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6313fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6314fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6315fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6316fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6317fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6318fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6319fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6320fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6321fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6322fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6323fa34dd3eSStefano Zampini if (zerodiag) { 6324fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6325fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 63264f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6327fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6328fa34dd3eSStefano Zampini } 6329fa34dd3eSStefano Zampini /* BDDC */ 6330fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6331fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6332fa34dd3eSStefano Zampini 6333fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6334fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6335fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6336fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 63374f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 63384f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6339fa34dd3eSStefano Zampini } 63404f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6341fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6342fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6343fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6344fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6345fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6346fa34dd3eSStefano Zampini } 6347fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6348fa34dd3eSStefano Zampini } 6349