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; 98672b8c272SStefano Zampini cum = 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 } 1000d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 100108122e43SStefano Zampini } 100208122e43SStefano Zampini 100308122e43SStefano Zampini if (mss) { /* multilevel */ 100408122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 100508122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 100608122e43SStefano Zampini } 100708122e43SStefano Zampini 1008ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 100908122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 101008122e43SStefano Zampini const PetscInt *idxs; 1011f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1012862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 101308122e43SStefano Zampini PetscBLASInt B_N; 1014aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 101508122e43SStefano Zampini 1016862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1017ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1018f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1019f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10209ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1021aff50787SStefano Zampini PetscInt j,k; 1022aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1023aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1024aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 102508122e43SStefano Zampini } 102608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1027aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1028aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1029aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1030aff50787SStefano Zampini } 103108122e43SStefano Zampini } 103208122e43SStefano Zampini } else { 103308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 103408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 103508122e43SStefano Zampini } 10368bec7fa6SStefano Zampini } else { 1037f6f667cfSStefano Zampini S = Sarray + cumarray; 1038f6f667cfSStefano Zampini St = Starray + cumarray; 10398bec7fa6SStefano Zampini } 1040aff50787SStefano Zampini /* see if we can save some work */ 1041aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1042aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1043aff50787SStefano Zampini } 1044aff50787SStefano Zampini 1045aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1046aff50787SStefano Zampini B_neigs = 0; 1047aff50787SStefano Zampini } else { 10489ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 104908122e43SStefano Zampini PetscBLASInt B_itype = 1; 1050f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 10514c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 10529552c7c7SStefano Zampini PetscInt nmin_s; 105308122e43SStefano Zampini 1054fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 10558bec7fa6SStefano 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]]); 1056fd14bc51SStefano Zampini } 1057d16cbb6bSStefano Zampini 105808122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1059d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1060d16cbb6bSStefano Zampini 1061d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 106208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1063f6f667cfSStefano 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)); 106408122e43SStefano Zampini #else 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,B_iwork,B_ifail,&B_ierr)); 106608122e43SStefano Zampini #endif 1067d16cbb6bSStefano Zampini } else { 1068d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1069d16cbb6bSStefano Zampini B_IL = 1; 1070d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1071d16cbb6bSStefano 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)); 1072d16cbb6bSStefano Zampini #else 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,B_iwork,B_ifail,&B_ierr)); 1074d16cbb6bSStefano Zampini #endif 1075d16cbb6bSStefano Zampini } 107608122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 107708122e43SStefano Zampini if (B_ierr) { 107808122e43SStefano Zampini if (B_ierr < 0 ) { 107908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 108008122e43SStefano Zampini } else if (B_ierr <= B_N) { 108108122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 108208122e43SStefano Zampini } else { 10839552c7c7SStefano 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); 108408122e43SStefano Zampini } 108508122e43SStefano Zampini } 108608122e43SStefano Zampini 108708122e43SStefano Zampini if (B_neigs > nmax) { 1088fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1089fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1090fd14bc51SStefano Zampini } 1091f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 109208122e43SStefano Zampini B_neigs = nmax; 109308122e43SStefano Zampini } 109408122e43SStefano Zampini 10959552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 10969552c7c7SStefano Zampini if (B_neigs < nmin_s) { 109708122e43SStefano Zampini PetscBLASInt B_neigs2; 109808122e43SStefano Zampini 1099f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1100f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1101fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1102fd14bc51SStefano 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); 1103fd14bc51SStefano Zampini } 11049ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 110508122e43SStefano Zampini PetscInt j; 110608122e43SStefano Zampini for (j=0;j<subset_size;j++) { 110708122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 110808122e43SStefano Zampini } 110908122e43SStefano Zampini for (j=0;j<subset_size;j++) { 111008122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 111108122e43SStefano Zampini } 111208122e43SStefano Zampini } else { 111308122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111408122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111508122e43SStefano Zampini } 111608122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 111708122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1118f6f667cfSStefano 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)); 111908122e43SStefano Zampini #else 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,B_iwork,B_ifail,&B_ierr)); 112108122e43SStefano Zampini #endif 112208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 112308122e43SStefano Zampini B_neigs += B_neigs2; 112408122e43SStefano Zampini } 112508122e43SStefano Zampini if (B_ierr) { 112608122e43SStefano Zampini if (B_ierr < 0 ) { 112708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 112808122e43SStefano Zampini } else if (B_ierr <= B_N) { 112908122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 113008122e43SStefano Zampini } else { 11319552c7c7SStefano 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); 113208122e43SStefano Zampini } 113308122e43SStefano Zampini } 1134fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1135ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 113608122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 113708122e43SStefano Zampini if (eigs[j] == 0.0) { 1138ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 113908122e43SStefano Zampini } else { 1140ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1141fd14bc51SStefano Zampini } 114208122e43SStefano Zampini } 114308122e43SStefano Zampini } 114408122e43SStefano Zampini } else { 114508122e43SStefano Zampini /* TODO */ 114608122e43SStefano Zampini } 1147aff50787SStefano Zampini } 11486c3e6151SStefano Zampini /* change the basis back to the original one */ 11496c3e6151SStefano Zampini if (sub_schurs->change) { 115072b8c272SStefano Zampini Mat change,phi,phit; 11516c3e6151SStefano Zampini 11526c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 11536c3e6151SStefano Zampini PetscInt ii; 11546c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 11556c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11566c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 11576c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 11586c3e6151SStefano Zampini } 11596c3e6151SStefano Zampini } 11606c3e6151SStefano Zampini } 116172b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 11626c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 116372b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 11646c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 11656c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 11666c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 11676c3e6151SStefano Zampini } 11688bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 11698bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 11709162d606SStefano Zampini if (B_neigs) { 11719162d606SStefano 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); 1172fd14bc51SStefano Zampini 1173fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 11749552c7c7SStefano Zampini PetscInt ii; 11759552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1176ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 11779552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1178ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1179ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1180ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1181ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1182ac47001eSStefano Zampini #else 1183ac47001eSStefano 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); 1184ac47001eSStefano Zampini #endif 11859552c7c7SStefano Zampini } 11869552c7c7SStefano Zampini } 1187fd14bc51SStefano Zampini } 11889162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 11899162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 11909162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 11919162d606SStefano Zampini cum++; 119208122e43SStefano Zampini } 119308122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 119408122e43SStefano Zampini /* shift for next computation */ 119508122e43SStefano Zampini cumarray += subset_size*subset_size; 119608122e43SStefano Zampini } 1197fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1198fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1199fd14bc51SStefano Zampini } 120008122e43SStefano Zampini 120108122e43SStefano Zampini if (mss) { 120208122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 120308122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1204f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1205f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1206f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 120708122e43SStefano Zampini } 1208f6f667cfSStefano Zampini if (allocated_S_St) { 1209f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1210f6f667cfSStefano Zampini } 1211f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 121208122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 121308122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 121408122e43SStefano Zampini #endif 121508122e43SStefano Zampini if (pcbddc->dbg_flag) { 12161b968477SStefano Zampini PetscInt maxneigs_r; 121708122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12189b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 121908122e43SStefano Zampini } 122008122e43SStefano Zampini PetscFunctionReturn(0); 122108122e43SStefano Zampini } 1222b1b3d7a2SStefano Zampini 1223674ae819SStefano Zampini #undef __FUNCT__ 1224c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1225c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1226c8587f34SStefano Zampini { 1227c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 12288629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1229c8587f34SStefano Zampini PetscErrorCode ierr; 1230c8587f34SStefano Zampini 1231c8587f34SStefano Zampini PetscFunctionBegin; 1232f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 12335e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1234c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1235c8587f34SStefano Zampini 1236684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 12370fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1238684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1239c8587f34SStefano Zampini 1240c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1241b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1242c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1243c8587f34SStefano Zampini } 1244c8587f34SStefano Zampini 12458629588bSStefano Zampini /* 12468629588bSStefano Zampini Setup local correction and local part of coarse basis. 12478629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 12488629588bSStefano Zampini */ 124947f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 12508629588bSStefano Zampini 12518629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 12528629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 12538629588bSStefano Zampini 12548629588bSStefano Zampini /* free */ 12558629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1256c8587f34SStefano Zampini PetscFunctionReturn(0); 1257c8587f34SStefano Zampini } 1258c8587f34SStefano Zampini 1259c8587f34SStefano Zampini #undef __FUNCT__ 1260674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1261674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1262674ae819SStefano Zampini { 1263674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1264674ae819SStefano Zampini PetscErrorCode ierr; 1265674ae819SStefano Zampini 1266674ae819SStefano Zampini PetscFunctionBegin; 1267674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1268674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 126930368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1270674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1271674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1272785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1273674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1274f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1275f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1276785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 127763602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 127863602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1279674ae819SStefano Zampini PetscFunctionReturn(0); 1280674ae819SStefano Zampini } 1281674ae819SStefano Zampini 1282674ae819SStefano Zampini #undef __FUNCT__ 1283674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1284674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1285674ae819SStefano Zampini { 1286674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 12874f1b2e48SStefano Zampini PetscInt i; 1288674ae819SStefano Zampini PetscErrorCode ierr; 1289674ae819SStefano Zampini 1290674ae819SStefano Zampini PetscFunctionBegin; 1291b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1292674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1293674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1294674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 12954f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 12964f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 12974f1b2e48SStefano Zampini } 12984f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1299b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1300674ae819SStefano Zampini PetscFunctionReturn(0); 1301674ae819SStefano Zampini } 1302674ae819SStefano Zampini 1303674ae819SStefano Zampini #undef __FUNCT__ 1304674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1305674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1306674ae819SStefano Zampini { 1307674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1308674ae819SStefano Zampini PetscErrorCode ierr; 1309674ae819SStefano Zampini 1310674ae819SStefano Zampini PetscFunctionBegin; 1311674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 131258da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1313ca92afb2SStefano Zampini PetscScalar *array; 131406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 131506656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 131658da7f69SStefano Zampini } 1317674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1318674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 131915aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 132015aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1321674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1322674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1323674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 132406656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1325674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1326674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13278ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1328674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1329674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1330674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1331f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1332f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1333f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1334f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1335727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 13360e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1337f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 133870cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 13396e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 134081d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 13410369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 13428b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 13434f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 13448b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1345ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1346ca92afb2SStefano Zampini PetscInt i; 1347ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1348ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1349ca92afb2SStefano Zampini } 1350ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1351ca92afb2SStefano Zampini } 13524f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1353674ae819SStefano Zampini PetscFunctionReturn(0); 1354674ae819SStefano Zampini } 1355674ae819SStefano Zampini 1356674ae819SStefano Zampini #undef __FUNCT__ 1357f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1358f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 13596bfb1811SStefano Zampini { 13606bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13616bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 13626bfb1811SStefano Zampini VecType impVecType; 13634f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 13646bfb1811SStefano Zampini PetscErrorCode ierr; 13656bfb1811SStefano Zampini 13666bfb1811SStefano Zampini PetscFunctionBegin; 1367f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1368019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1369f4ddd8eeSStefano Zampini } 1370e7b262bdSStefano Zampini /* get sizes */ 13714f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1372b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 13736bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1374e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1375e7b262bdSStefano Zampini /* R nodes */ 1376e7b262bdSStefano Zampini old_size = -1; 1377e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1378e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1379e7b262bdSStefano Zampini } 1380e7b262bdSStefano Zampini if (n_R != old_size) { 1381e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1382e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 13836bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 13846bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 13856bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 13866bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1387e7b262bdSStefano Zampini } 1388e7b262bdSStefano Zampini /* local primal dofs */ 1389e7b262bdSStefano Zampini old_size = -1; 1390e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1391e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1392e7b262bdSStefano Zampini } 1393e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1394e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 139583b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1396e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 13976bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1398e7b262bdSStefano Zampini } 1399e7b262bdSStefano Zampini /* local explicit constraints */ 1400e7b262bdSStefano Zampini old_size = -1; 1401e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1402e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1403e7b262bdSStefano Zampini } 1404e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1405e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 140683b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 140783b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 140883b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 140983b7ccabSStefano Zampini } 14106bfb1811SStefano Zampini PetscFunctionReturn(0); 14116bfb1811SStefano Zampini } 14126bfb1811SStefano Zampini 14136bfb1811SStefano Zampini #undef __FUNCT__ 141447f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 141547f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 141688ebb749SStefano Zampini { 141725084f0cSStefano Zampini PetscErrorCode ierr; 141825084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 141988ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 142088ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1421d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 142225084f0cSStefano Zampini /* submatrices of local problem */ 142380677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 142406656605SStefano Zampini /* submatrices of local coarse problem */ 142506656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 142625084f0cSStefano Zampini /* working matrices */ 142706656605SStefano Zampini Mat C_CR; 142825084f0cSStefano Zampini /* additional working stuff */ 142906656605SStefano Zampini PC pc_R; 14304f1b2e48SStefano Zampini Mat F; 1431a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 143206656605SStefano Zampini 143325084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 143406656605SStefano Zampini PetscScalar *work; 143506656605SStefano Zampini PetscInt *idx_V_B; 1436ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 143706656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1438ffd830a3SStefano Zampini 143925084f0cSStefano Zampini /* some shortcuts to scalars */ 144006656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 144188ebb749SStefano Zampini 144288ebb749SStefano Zampini PetscFunctionBegin; 1443ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1444ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1445ffd830a3SStefano Zampini } 1446ffd830a3SStefano Zampini 1447ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1448b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 14494f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1450b371cd4fSStefano Zampini n_B = pcis->n_B; 1451b371cd4fSStefano Zampini n_D = pcis->n - n_B; 145288ebb749SStefano Zampini n_R = pcis->n - n_vertices; 145388ebb749SStefano Zampini 145488ebb749SStefano Zampini /* vertices in boundary numbering */ 1455785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 14560e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 145788ebb749SStefano Zampini if (i != n_vertices) { 145822d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 145988ebb749SStefano Zampini } 146088ebb749SStefano Zampini 146106656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1462019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 146306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 146406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 146506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 146606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 146706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 146806656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 146906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 147006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 147106656605SStefano Zampini 147206656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 147306656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 147406656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 147506656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 147606656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1477ffd830a3SStefano Zampini lda_rhs = n_R; 1478a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 147906656605SStefano Zampini if (isLU || isILU || isCHOL) { 148006656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1481df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1482df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1483d62866d3SStefano Zampini MatFactorType type; 1484d62866d3SStefano Zampini 1485df4d28bfSStefano Zampini F = reuse_solver->F; 14866816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1487d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1488ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1489df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 149006656605SStefano Zampini } else { 149106656605SStefano Zampini F = NULL; 149206656605SStefano Zampini } 149306656605SStefano Zampini 1494ffd830a3SStefano Zampini /* allocate workspace */ 1495ffd830a3SStefano Zampini n = 0; 1496ffd830a3SStefano Zampini if (n_constraints) { 1497ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1498ffd830a3SStefano Zampini } 1499ffd830a3SStefano Zampini if (n_vertices) { 1500ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1501ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1502ffd830a3SStefano Zampini } 1503ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1504ffd830a3SStefano Zampini 150588ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 150688ebb749SStefano Zampini if (n_constraints) { 150772b8c272SStefano Zampini Mat M1,M2,M3,C_B; 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); 151772b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);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) { 1539df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1540a3df083aSStefano Zampini 154172b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 154272b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1543a3df083aSStefano Zampini } 154480677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1545a3df083aSStefano Zampini if (need_benign_correction) { 1546a3df083aSStefano Zampini PetscScalar *marr; 1547df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1548a3df083aSStefano Zampini 1549a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1550a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1551a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1552df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1553a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1554a3df083aSStefano Zampini } 1555a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1556a3df083aSStefano Zampini } 155706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 155806656605SStefano Zampini } else { 155980677318SStefano Zampini PetscScalar *marr; 156080677318SStefano Zampini 156180677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 156206656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1563ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1564ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 156506656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 156606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 156706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 156806656605SStefano Zampini } 156980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 157006656605SStefano Zampini } 157180677318SStefano Zampini if (!pcbddc->switch_static) { 157280677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 157380677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 157480677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 157580677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1576ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 157780677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 157880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 157980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 158080677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 158180677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 158280677318SStefano Zampini } 158380677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 158480677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 158572b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 158680677318SStefano Zampini } else { 1587ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1588ffd830a3SStefano Zampini IS dummy; 1589ffd830a3SStefano Zampini 1590ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 159172b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1592ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1593ffd830a3SStefano Zampini } else { 159480677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 159580677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1596ffd830a3SStefano Zampini } 159725084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 159880677318SStefano Zampini } 159980677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 160080677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 160180677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 160206656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 160306656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 160480677318SStefano Zampini if (isCHOL) { 160580677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 160680677318SStefano Zampini } else { 160725084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 160880677318SStefano Zampini } 160980677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 161006656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 161125084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 161225084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 161325084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 161480677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 161572b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 161672b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 161706656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 161806656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1619f4ddd8eeSStefano Zampini } 1620fc227af8SStefano Zampini 1621fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 162288ebb749SStefano Zampini if (n_vertices) { 162306656605SStefano Zampini IS is_aux; 16243a50541eSStefano Zampini 1625df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 16266816873aSStefano Zampini IS tis; 16276816873aSStefano Zampini 16286816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 16296816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 16306816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 16316816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 16326816873aSStefano Zampini } else { 16333a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 16346816873aSStefano Zampini } 16359577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 16369577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 163704708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 163825084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 163988ebb749SStefano Zampini } 164088ebb749SStefano Zampini 164188ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1642f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 164306656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 164406656605SStefano Zampini if (pcbddc->coarse_phi_D) { 164506656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 164606656605SStefano Zampini } 1647f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 164806656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 164906656605SStefano Zampini PetscScalar *marray; 165006656605SStefano Zampini 165106656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 165206656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1653f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1654f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1655f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1656f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1657f4ddd8eeSStefano Zampini } 1658f4ddd8eeSStefano Zampini } 165906656605SStefano Zampini 1660f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 166106656605SStefano Zampini PetscScalar *marray; 166288ebb749SStefano Zampini 166306656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16648eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 166506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 166688ebb749SStefano Zampini } 16673301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 166806656605SStefano Zampini n *= 2; 166988ebb749SStefano Zampini } 167006656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 167106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 167206656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 16738eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 167406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 167506656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 167688ebb749SStefano Zampini } 16773301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 167806656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 16798eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 168006656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 168106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 168288ebb749SStefano Zampini } 168388ebb749SStefano Zampini } else { 1684c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1685c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 16861b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1687c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1688c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1689c0553b1fSStefano Zampini } 169088ebb749SStefano Zampini } 169106656605SStefano Zampini } 1692019a44ceSStefano Zampini 169306656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 16944f1b2e48SStefano Zampini p0_lidx_I = NULL; 16954f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1696d12edf2fSStefano Zampini const PetscInt *idxs; 1697d12edf2fSStefano Zampini 1698d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 16994f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17004f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17014f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17024f1b2e48SStefano Zampini } 1703d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1704d12edf2fSStefano Zampini } 1705d16cbb6bSStefano Zampini 170606656605SStefano Zampini /* vertices */ 170706656605SStefano Zampini if (n_vertices) { 170816f15bc4SStefano Zampini 1709ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 171004708bb6SStefano Zampini 171116f15bc4SStefano Zampini if (n_R) { 171214393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 171306656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 171416f15bc4SStefano Zampini PetscScalar *x,*y; 171504708bb6SStefano Zampini PetscBool isseqaij; 171606656605SStefano Zampini 171721eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 171814393ed6SStefano Zampini if (need_benign_correction) { 171914393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 172014393ed6SStefano Zampini IS is_p0; 172114393ed6SStefano Zampini PetscInt *idxs_p0,n; 172214393ed6SStefano Zampini 172314393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 172414393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 172514393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 172614393ed6SStefano Zampini if (n != pcbddc->benign_n) { 172714393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 172814393ed6SStefano Zampini } 172914393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 173014393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 173114393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 173214393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 173314393ed6SStefano Zampini } 173414393ed6SStefano Zampini 1735ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1736ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1737ffd830a3SStefano Zampini } else { 1738ca92afb2SStefano Zampini PetscScalar *av,*array; 1739ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1740ca92afb2SStefano Zampini PetscInt n; 1741ca92afb2SStefano Zampini PetscBool flg_row; 1742ffd830a3SStefano Zampini 1743ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1744ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1745ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1746ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1747ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1748ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1749ca92afb2SStefano Zampini PetscInt j; 1750ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1751ffd830a3SStefano Zampini } 1752ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1753ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1754ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1755ffd830a3SStefano Zampini } 1756ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1757a3df083aSStefano Zampini if (need_benign_correction) { 1758df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1759a3df083aSStefano Zampini PetscScalar *marr; 1760a3df083aSStefano Zampini 1761a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 176214393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 176314393ed6SStefano Zampini 176414393ed6SStefano Zampini | 0 0 0 | (V) 176514393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 176614393ed6SStefano Zampini | 0 0 -1 | (p0) 176714393ed6SStefano Zampini 176814393ed6SStefano Zampini */ 1769df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 177014393ed6SStefano Zampini const PetscScalar *vals; 177114393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 177214393ed6SStefano Zampini PetscInt n,j,nz; 177314393ed6SStefano Zampini 1774df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1775df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 177614393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 177714393ed6SStefano Zampini for (j=0;j<n;j++) { 177814393ed6SStefano Zampini PetscScalar val = vals[j]; 177914393ed6SStefano Zampini PetscInt k,col = idxs[j]; 178014393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 178114393ed6SStefano Zampini } 178214393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1783df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 178414393ed6SStefano Zampini } 178572b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 178672b8c272SStefano Zampini } 178772b8c272SStefano Zampini if (F) { 178814393ed6SStefano Zampini /* need to correct the rhs */ 178972b8c272SStefano Zampini if (need_benign_correction) { 179072b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 179172b8c272SStefano Zampini PetscScalar *marr; 179272b8c272SStefano Zampini 179372b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 1794a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1795a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1796df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1797a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1798a3df083aSStefano Zampini } 1799a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1800a3df083aSStefano Zampini } 180106656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 180214393ed6SStefano Zampini /* need to correct the solution */ 1803a3df083aSStefano Zampini if (need_benign_correction) { 1804df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1805a3df083aSStefano Zampini PetscScalar *marr; 1806a3df083aSStefano Zampini 1807a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1808a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1809a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1810df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1811a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1812a3df083aSStefano Zampini } 1813a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1814a3df083aSStefano Zampini } 181506656605SStefano Zampini } else { 181606656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 181706656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1818ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1819ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 182006656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 182106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 182206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 182306656605SStefano Zampini } 182406656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 182506656605SStefano Zampini } 182680677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1827ffd830a3SStefano Zampini /* S_VV and S_CV */ 182806656605SStefano Zampini if (n_constraints) { 182906656605SStefano Zampini Mat B; 183080677318SStefano Zampini 1831ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 183280677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1833ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1834ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 183580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 183680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 183780677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 183880677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 183980677318SStefano Zampini } 1840ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 184180677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 184280677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1843ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 184480677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 184506656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1846ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1847ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 184806656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 184906656605SStefano Zampini } 185004708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 185104708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 185204708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 185304708bb6SStefano Zampini } 1854ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1855ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1856ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1857ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1858ffd830a3SStefano Zampini } 185906656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 186014393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 186114393ed6SStefano Zampini if (need_benign_correction) { 1862df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 186314393ed6SStefano Zampini PetscScalar *marr,*sums; 186414393ed6SStefano Zampini 186514393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 186614393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1867df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 186814393ed6SStefano Zampini const PetscScalar *vals; 186914393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 187014393ed6SStefano Zampini PetscInt n,j,nz; 187114393ed6SStefano Zampini 1872df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1873df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 187414393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 187514393ed6SStefano Zampini PetscInt k; 187614393ed6SStefano Zampini sums[j] = 0.; 187714393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 187814393ed6SStefano Zampini } 187914393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 188014393ed6SStefano Zampini for (j=0;j<n;j++) { 188114393ed6SStefano Zampini PetscScalar val = vals[j]; 188214393ed6SStefano Zampini PetscInt k; 188314393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 188414393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 188514393ed6SStefano Zampini } 188614393ed6SStefano Zampini } 188714393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1888df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 188914393ed6SStefano Zampini } 189014393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 189114393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 189214393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 189314393ed6SStefano Zampini } 189480677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 189506656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 189606656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 189706656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 189806656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 189906656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 190006656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 190106656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1902d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1903019a44ceSStefano Zampini } else { 1904d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1905d16cbb6bSStefano Zampini } 190621eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1907d16cbb6bSStefano Zampini 190806656605SStefano Zampini /* coarse basis functions */ 190906656605SStefano Zampini for (i=0;i<n_vertices;i++) { 191016f15bc4SStefano Zampini PetscScalar *y; 191116f15bc4SStefano Zampini 1912ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 191306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 191406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 191506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 191606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 191706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 191806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 191906656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 192006656605SStefano Zampini 192106656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19224f1b2e48SStefano Zampini PetscInt j; 19234f1b2e48SStefano Zampini 192406656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 192506656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 192606656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192706656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 192806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19294f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 193006656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 193106656605SStefano Zampini } 193206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 193306656605SStefano Zampini } 193404708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 193504708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 193606656605SStefano Zampini } 193706656605SStefano Zampini 193806656605SStefano Zampini if (n_constraints) { 193906656605SStefano Zampini Mat B; 194006656605SStefano Zampini 1941ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 194206656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 194380677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 194406656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 194506656605SStefano Zampini if (n_vertices) { 194680677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 194780677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 194880677318SStefano Zampini } else { 194980677318SStefano Zampini Mat S_VCt; 195080677318SStefano Zampini 1951ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1952ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 195372b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 1954ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 1955ffd830a3SStefano Zampini } 195680677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 195780677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 195880677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 195980677318SStefano Zampini } 196006656605SStefano Zampini } 196106656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 196206656605SStefano Zampini /* coarse basis functions */ 196306656605SStefano Zampini for (i=0;i<n_constraints;i++) { 196406656605SStefano Zampini PetscScalar *y; 196506656605SStefano Zampini 1966ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 196706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 196806656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 196906656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197006656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197106656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 197206656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 197306656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19744f1b2e48SStefano Zampini PetscInt j; 19754f1b2e48SStefano Zampini 197606656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 197706656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 197806656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 197906656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 198006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 19814f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 198206656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 198306656605SStefano Zampini } 198406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 198506656605SStefano Zampini } 198606656605SStefano Zampini } 198780677318SStefano Zampini if (n_constraints) { 198880677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 198980677318SStefano Zampini } 19904f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 199172b8c272SStefano Zampini 199272b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 199372b8c272SStefano Zampini if (pcbddc->benign_n) { 199472b8c272SStefano Zampini Mat B0_B,B0_BPHI; 199572b8c272SStefano Zampini IS is_dummy; 199672b8c272SStefano Zampini PetscScalar *data; 199772b8c272SStefano Zampini PetscInt j; 199872b8c272SStefano Zampini 199972b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 200072b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 200172b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 200272b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 200372b8c272SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 200472b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 200572b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 200672b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 200772b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 200872b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 200972b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 201072b8c272SStefano Zampini } 201172b8c272SStefano Zampini } 201272b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 201372b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 201472b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 201572b8c272SStefano Zampini } 2016019a44ceSStefano Zampini 201706656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 20183301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2019ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2020ffd830a3SStefano Zampini PetscScalar *marray; 202106656605SStefano Zampini 202206656605SStefano Zampini if (n_constraints) { 2023ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 202406656605SStefano Zampini 2025ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 202606656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2027ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 202816f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 202906656605SStefano Zampini if (n_vertices) { 2030ffd830a3SStefano Zampini Mat S_VCT; 203106656605SStefano Zampini 203206656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2033ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 203416f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 203506656605SStefano Zampini } 2036ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 203706656605SStefano Zampini } 203816f15bc4SStefano Zampini if (n_vertices && n_R) { 2039ffd830a3SStefano Zampini PetscScalar *av,*marray; 2040ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2041ffd830a3SStefano Zampini PetscInt n; 2042ffd830a3SStefano Zampini PetscBool flg_row; 204306656605SStefano Zampini 2044ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2045ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2046ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2047ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2048ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2049ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2050ffd830a3SStefano Zampini PetscInt j; 2051ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2052ffd830a3SStefano Zampini } 2053ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2054ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2055ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 205606656605SStefano Zampini } 205706656605SStefano Zampini 2058ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2059ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2060ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2061ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2062ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 206306656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 206406656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 206506656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 206606656605SStefano Zampini } 2067ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2068ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2069ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2070ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2071ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2072ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2073ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2074ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 207506656605SStefano Zampini } 2076ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 207706656605SStefano Zampini /* coarse basis functions */ 207806656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 207906656605SStefano Zampini PetscScalar *y; 208006656605SStefano Zampini 2081ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 208206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 208306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 208406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 208606656605SStefano Zampini if (i<n_vertices) { 208706656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 208806656605SStefano Zampini } 208906656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 209006656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 209106656605SStefano Zampini 209206656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 209306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 209406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 209506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 209606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 209706656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 209806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 209906656605SStefano Zampini } 210006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 210106656605SStefano Zampini } 2102ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2103ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 210406656605SStefano Zampini } 2105d62866d3SStefano Zampini /* free memory */ 210688ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 210706656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 210806656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 210906656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 211006656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2111d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2112d62866d3SStefano Zampini if (n_vertices) { 2113d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2114d62866d3SStefano Zampini } 2115d62866d3SStefano Zampini if (n_constraints) { 2116d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2117d62866d3SStefano Zampini } 211888ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 211988ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 212088ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2121d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 212288ebb749SStefano Zampini Mat coarse_sub_mat; 212325084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 212488ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 212588ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 212688ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 21278bec7fa6SStefano Zampini Mat C_B,CPHI; 21288bec7fa6SStefano Zampini IS is_dummy; 21298bec7fa6SStefano Zampini Vec mones; 213088ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 213188ebb749SStefano Zampini PetscReal real_value; 213288ebb749SStefano Zampini 2133a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2134a3df083aSStefano Zampini Mat A; 2135a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2136a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2137a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2138a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2139a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2140a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2141a3df083aSStefano Zampini } else { 214288ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 214388ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 214488ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 214588ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2146a3df083aSStefano Zampini } 214788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 214888ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2149ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 215088ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 215188ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 215288ebb749SStefano Zampini } 215388ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 215488ebb749SStefano Zampini 215525084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 21563301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 215725084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2158ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 215988ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 216088ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 216188ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216288ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 216388ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 216488ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216588ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 216688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 216788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 216888ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 216988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 217088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 217188ebb749SStefano Zampini } else { 217288ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 217388ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 217488ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 217588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 217688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 217788ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 217888ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 217988ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 218088ebb749SStefano Zampini } 218188ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 218288ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 218388ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 218488ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 21854f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2186fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2187d12edf2fSStefano Zampini PetscScalar *data,*data2; 21884f1b2e48SStefano Zampini PetscInt j; 2189d12edf2fSStefano Zampini 21904f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2191fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2192d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2193d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2194d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2195d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 21964f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 21974f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2198d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 21994f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 22004f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22014f1b2e48SStefano Zampini } 2202d12edf2fSStefano Zampini } 2203d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2204d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2205d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2206d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2207d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2208d12edf2fSStefano Zampini } 2209d12edf2fSStefano Zampini #if 0 2210d12edf2fSStefano Zampini { 2211d12edf2fSStefano Zampini PetscViewer viewer; 2212d12edf2fSStefano Zampini char filename[256]; 2213ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2214d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2215d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2216ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2217ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2218ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2219d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 222072b8c272SStefano Zampini if (save_change) { 222172b8c272SStefano Zampini Mat phi_B; 222272b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 222372b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 222472b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 222572b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 222672b8c272SStefano Zampini } else { 2227ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2228ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 222972b8c272SStefano Zampini } 2230ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2231ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2232ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2233ffd830a3SStefano Zampini } 2234ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2235ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2236ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2237ffd830a3SStefano Zampini } 223872b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2239ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2240ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2241ffd830a3SStefano Zampini } 2242d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2243d12edf2fSStefano Zampini } 2244d12edf2fSStefano Zampini #endif 224581d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 22468bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 22471575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 224806656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 22498bec7fa6SStefano Zampini 22508bec7fa6SStefano Zampini /* check constraints */ 22514f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 22528bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 22538bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 22548bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 22558bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 22568bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 22578bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 22588bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2259bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2260ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2261bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2262bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2263bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2264bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2265bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 226688ebb749SStefano Zampini } 22678bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 22688bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 22698bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 22708bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2271d12edf2fSStefano Zampini } 227225084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 227388ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 227488ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 227588ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 227688ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 227788ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 227888ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 227988ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 228088ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 228188ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 228288ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2283ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 228488ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 228588ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 228688ebb749SStefano Zampini } 228788ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 228888ebb749SStefano Zampini } 22898629588bSStefano Zampini /* get back data */ 22908629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 229188ebb749SStefano Zampini PetscFunctionReturn(0); 229288ebb749SStefano Zampini } 229388ebb749SStefano Zampini 229488ebb749SStefano Zampini #undef __FUNCT__ 2295d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2296d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2297aa0d41d4SStefano Zampini { 2298d65f70fdSStefano Zampini Mat *work_mat; 2299d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2300d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2301d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2302aa0d41d4SStefano Zampini PetscErrorCode ierr; 2303aa0d41d4SStefano Zampini 2304aa0d41d4SStefano Zampini PetscFunctionBegin; 2305d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2306d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2307d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2308d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2309aa0d41d4SStefano Zampini 2310d65f70fdSStefano Zampini if (!rsorted) { 2311906d46d4SStefano Zampini const PetscInt *idxs; 2312906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2313aa0d41d4SStefano Zampini 2314d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2315d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2316d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2317d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2318aa0d41d4SStefano Zampini } 2319d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2320d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2321d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2322d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2323aa0d41d4SStefano Zampini } 2324d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2325d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2326d65f70fdSStefano Zampini } else { 2327d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2328d65f70fdSStefano Zampini isrow_s = isrow; 2329aa0d41d4SStefano Zampini } 2330906d46d4SStefano Zampini 2331d65f70fdSStefano Zampini if (!csorted) { 2332d65f70fdSStefano Zampini if (isrow == iscol) { 2333d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2334d65f70fdSStefano Zampini iscol_s = isrow_s; 2335d65f70fdSStefano Zampini } else { 2336d65f70fdSStefano Zampini const PetscInt *idxs; 2337d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2338906d46d4SStefano Zampini 2339d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2340d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2341d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2342d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2343d65f70fdSStefano Zampini } 2344d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2345d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2346d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2347d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2348d65f70fdSStefano Zampini } 2349d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2350d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2351d65f70fdSStefano Zampini } 2352d65f70fdSStefano Zampini } else { 2353d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2354d65f70fdSStefano Zampini iscol_s = iscol; 2355d65f70fdSStefano Zampini } 2356d65f70fdSStefano Zampini 2357d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2358d65f70fdSStefano Zampini 2359d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2360906d46d4SStefano Zampini Mat new_mat; 2361d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2362906d46d4SStefano Zampini 2363d65f70fdSStefano Zampini if (!rsorted) { 2364d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2365d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2366d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2367d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2368906d46d4SStefano Zampini } 2369d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2370d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2371d65f70fdSStefano Zampini } else { 2372d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2373906d46d4SStefano Zampini } 2374d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2375d65f70fdSStefano Zampini 2376d65f70fdSStefano Zampini if (!csorted) { 2377d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2378d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2379d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2380d65f70fdSStefano Zampini } else { 2381d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2382d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2383d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2384d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2385d65f70fdSStefano Zampini } 2386d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2387d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2388d65f70fdSStefano Zampini } 2389d65f70fdSStefano Zampini } else { 2390d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2391d65f70fdSStefano Zampini } 2392d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2393d65f70fdSStefano Zampini 2394d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2395d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2396d65f70fdSStefano Zampini work_mat[0] = new_mat; 2397d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2398d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2399d65f70fdSStefano Zampini } 2400d65f70fdSStefano Zampini 2401d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2402d65f70fdSStefano Zampini *B = work_mat[0]; 2403d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2404d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2405d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2406d65f70fdSStefano Zampini PetscFunctionReturn(0); 2407d65f70fdSStefano Zampini } 2408d65f70fdSStefano Zampini 2409d65f70fdSStefano Zampini #undef __FUNCT__ 24105e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 24115e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2412aa0d41d4SStefano Zampini { 2413aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 24145e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2415d65f70fdSStefano Zampini Mat new_mat; 24165e8657edSStefano Zampini IS is_local,is_global; 2417d65f70fdSStefano Zampini PetscInt local_size; 2418d65f70fdSStefano Zampini PetscBool isseqaij; 2419aa0d41d4SStefano Zampini PetscErrorCode ierr; 2420aa0d41d4SStefano Zampini 2421aa0d41d4SStefano Zampini PetscFunctionBegin; 2422aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 24235e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 24245e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2425b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2426aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2427d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2428aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2429906d46d4SStefano Zampini 2430906d46d4SStefano Zampini /* check */ 2431906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2432906d46d4SStefano Zampini Vec x,x_change; 2433906d46d4SStefano Zampini PetscReal error; 2434906d46d4SStefano Zampini 24355e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2436906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 24375e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2438e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2439e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2440d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2441e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2442e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2443906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2444906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2445906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2446906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2447906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2448906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2449906d46d4SStefano Zampini } 2450906d46d4SStefano Zampini 245122d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 24529b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 245322d5777bSStefano Zampini if (isseqaij) { 24541cf9b237SStefano Zampini Mat M; 24551cf9b237SStefano Zampini 24561cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24571cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24581cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2459aa0d41d4SStefano Zampini } else { 24601cf9b237SStefano Zampini Mat work_mat,M; 24611cf9b237SStefano Zampini 2462aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 24631cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 24641cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 24651cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2466aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2467aa0d41d4SStefano Zampini } 24683301b35fSStefano Zampini if (matis->A->symmetric_set) { 24693301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2470e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 24713301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2472e496cd5dSStefano Zampini #endif 24733301b35fSStefano Zampini } 247445a1bb75SStefano Zampini /* 247545a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2476d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 247745a1bb75SStefano Zampini */ 2478d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2479aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2480aa0d41d4SStefano Zampini } 2481aa0d41d4SStefano Zampini 2482aa0d41d4SStefano Zampini #undef __FUNCT__ 2483a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 24848ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2485a64d13efSStefano Zampini { 2486a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2487a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2488d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 248953892102SStefano Zampini PetscInt *idx_R_local=NULL; 24903a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 24913a50541eSStefano Zampini PetscInt vbs,bs; 24926816873aSStefano Zampini PetscBT bitmask=NULL; 2493a64d13efSStefano Zampini PetscErrorCode ierr; 2494a64d13efSStefano Zampini 2495a64d13efSStefano Zampini PetscFunctionBegin; 2496b23d619eSStefano Zampini /* 2497b23d619eSStefano Zampini No need to setup local scatters if 2498b23d619eSStefano Zampini - primal space is unchanged 2499b23d619eSStefano Zampini AND 2500b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2501b23d619eSStefano Zampini AND 2502b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2503b23d619eSStefano Zampini */ 2504b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2505f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2506f4ddd8eeSStefano Zampini } 2507f4ddd8eeSStefano Zampini /* destroy old objects */ 2508f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2509f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2510f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2511a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2512b371cd4fSStefano Zampini n_B = pcis->n_B; 2513b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2514b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 25153a50541eSStefano Zampini 2516a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 25176816873aSStefano Zampini 251853892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2519df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2520854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2521a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2522a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25230e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2524a64d13efSStefano Zampini } 2525a64d13efSStefano Zampini 2526a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 25274641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 25286816873aSStefano Zampini idx_R_local[n_R++] = i; 2529a64d13efSStefano Zampini } 2530a64d13efSStefano Zampini } 2531df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2532df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 25336816873aSStefano Zampini 2534df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2535df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 25366816873aSStefano Zampini } 25373a50541eSStefano Zampini 25383a50541eSStefano Zampini /* Block code */ 25393a50541eSStefano Zampini vbs = 1; 25403a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 25413a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 25423a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 25433a50541eSStefano Zampini PetscInt *vary; 2544df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2545785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 25463a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2547d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2548d3df7717SStefano 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 */ 25490e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2550d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 25513a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 25523a50541eSStefano Zampini is_blocked = PETSC_FALSE; 25533a50541eSStefano Zampini break; 25543a50541eSStefano Zampini } 25553a50541eSStefano Zampini } 2556d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2557d3df7717SStefano Zampini } else { 2558d3df7717SStefano Zampini /* Verify directly the R set */ 2559d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2560d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2561d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2562d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2563d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2564d3df7717SStefano Zampini break; 2565d3df7717SStefano Zampini } 2566d3df7717SStefano Zampini } 2567d3df7717SStefano Zampini } 2568d3df7717SStefano Zampini } 25693a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 25703a50541eSStefano Zampini vbs = bs; 25713a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 25723a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 25733a50541eSStefano Zampini } 25743a50541eSStefano Zampini } 25753a50541eSStefano Zampini } 25763a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2577df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2578df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 257953892102SStefano Zampini 2580df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2581df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 258253892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2583df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 258453892102SStefano Zampini } else { 25853a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 258653892102SStefano Zampini } 2587a64d13efSStefano Zampini 2588a64d13efSStefano Zampini /* print some info if requested */ 2589a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2590a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2591a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 25921575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2593a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2594a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 25954f1b2e48SStefano 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); 2596a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2597a64d13efSStefano Zampini } 2598a64d13efSStefano Zampini 2599a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2600df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 26016816873aSStefano Zampini IS is_aux1,is_aux2; 26026816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 26036816873aSStefano Zampini 26043a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2605854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2606854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2607a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26084641a718SStefano Zampini for (i=0; i<n_D; i++) { 26094641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26104641a718SStefano Zampini } 2611a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2612a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26134641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 26144641a718SStefano Zampini aux_array1[j++] = i; 2615a64d13efSStefano Zampini } 2616a64d13efSStefano Zampini } 2617a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2618a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2619a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 26204641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 26214641a718SStefano Zampini aux_array2[j++] = i; 2622a64d13efSStefano Zampini } 2623a64d13efSStefano Zampini } 2624a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2625a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2626a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2627a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2628a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2629a64d13efSStefano Zampini 26308eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2631785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2632a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26334641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 26344641a718SStefano Zampini aux_array1[j++] = i; 2635a64d13efSStefano Zampini } 2636a64d13efSStefano Zampini } 2637a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2638a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2639a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2640a64d13efSStefano Zampini } 26414641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 26423a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2643d62866d3SStefano Zampini } else { 2644df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26456816873aSStefano Zampini IS tis; 26466816873aSStefano Zampini PetscInt schur_size; 26476816873aSStefano Zampini 2648df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 26496816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2650df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 26516816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 26526816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 26536816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 26546816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 26556816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2656d62866d3SStefano Zampini } 2657d62866d3SStefano Zampini } 2658a64d13efSStefano Zampini PetscFunctionReturn(0); 2659a64d13efSStefano Zampini } 2660a64d13efSStefano Zampini 2661304d26faSStefano Zampini 2662304d26faSStefano Zampini #undef __FUNCT__ 2663304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2664684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2665304d26faSStefano Zampini { 2666304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2667304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2668304d26faSStefano Zampini PC pc_temp; 2669304d26faSStefano Zampini Mat A_RR; 2670f4ddd8eeSStefano Zampini MatReuse reuse; 2671304d26faSStefano Zampini PetscScalar m_one = -1.0; 2672304d26faSStefano Zampini PetscReal value; 267304708bb6SStefano Zampini PetscInt n_D,n_R; 26749577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2675304d26faSStefano Zampini PetscErrorCode ierr; 2676e604994aSStefano Zampini /* prefixes stuff */ 2677312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2678e604994aSStefano Zampini size_t len; 2679304d26faSStefano Zampini 2680304d26faSStefano Zampini PetscFunctionBegin; 2681304d26faSStefano Zampini 2682e604994aSStefano Zampini /* compute prefixes */ 2683e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2684e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2685e604994aSStefano Zampini if (!pcbddc->current_level) { 2686e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2687e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2688e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2689e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2690e604994aSStefano Zampini } else { 2691e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2692312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2693e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2694e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2695312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2696312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 269734d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 269834d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2699e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2700e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2701e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2702e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2703e604994aSStefano Zampini } 2704e604994aSStefano Zampini 2705304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2706684f6988SStefano Zampini if (dirichlet) { 2707d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2708450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2709450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2710450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2711450f8f5eSStefano Zampini } 2712450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2713a3df083aSStefano Zampini Mat A_IIn; 2714a3df083aSStefano Zampini 2715a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2716a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2717a3df083aSStefano Zampini pcis->A_II = A_IIn; 2718a3df083aSStefano Zampini } 2719450f8f5eSStefano Zampini } 27203301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27213301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2722964fefecSStefano Zampini } 2723ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2724964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2725304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2726304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2727304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2728304d26faSStefano Zampini /* default */ 2729304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2730e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 27319577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2732304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 27339577ea80SStefano Zampini if (issbaij) { 27349577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 27359577ea80SStefano Zampini } else { 2736304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 27379577ea80SStefano Zampini } 2738304d26faSStefano Zampini /* Allow user's customization */ 2739304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2740304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2741304d26faSStefano Zampini } 2742d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2743df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2744df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2745d62866d3SStefano Zampini 2746df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2747d5574798SStefano Zampini } 2748304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2749304d26faSStefano Zampini if (!n_D) { 2750304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2751304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2752304d26faSStefano Zampini } 2753304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2754304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2755304d26faSStefano Zampini /* set ksp_D into pcis data */ 2756304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2757304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2758304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2759684f6988SStefano Zampini } 2760304d26faSStefano Zampini 2761304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2762684f6988SStefano Zampini A_RR = 0; 2763684f6988SStefano Zampini if (neumann) { 2764d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 276504708bb6SStefano Zampini PetscInt ibs,mbs; 276604708bb6SStefano Zampini PetscBool issbaij; 276704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2768f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 27698ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2770f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2771f4ddd8eeSStefano Zampini PetscInt nn_R; 277281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2773f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2774f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2775f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2776f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2777f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2778f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2779f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2780727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2781f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2782f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2783f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2784f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2785f4ddd8eeSStefano Zampini } 2786f4ddd8eeSStefano Zampini } 2787f4ddd8eeSStefano Zampini /* last check */ 2788d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2789f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2790f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2791f4ddd8eeSStefano Zampini } 2792f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2793f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2794f4ddd8eeSStefano Zampini } 2795f4ddd8eeSStefano Zampini /* extract A_RR */ 2796af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2797af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 279804708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 279904708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 280004708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 280104708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 280204708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2803af732b37SStefano Zampini } else { 280404708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 28056816873aSStefano Zampini } 280604708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 280704708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 280804708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 280904708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 281004708bb6SStefano Zampini } else { 281104708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 281204708bb6SStefano Zampini } 281304708bb6SStefano Zampini } 281416e386b8SStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 281516e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 281616e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 281716e386b8SStefano Zampini } else { 2818f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 281916e386b8SStefano Zampini } 28203301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28213301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 28226816873aSStefano Zampini } 2823f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2824304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2825304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2826304d26faSStefano Zampini /* default */ 2827304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2828e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2829304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 28309577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 28319577ea80SStefano Zampini if (issbaij) { 28329577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28339577ea80SStefano Zampini } else { 2834304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28359577ea80SStefano Zampini } 2836304d26faSStefano Zampini /* Allow user's customization */ 2837304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2838304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2839304d26faSStefano Zampini } 2840304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2841304d26faSStefano Zampini if (!n_R) { 2842304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2843304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2844304d26faSStefano Zampini } 2845df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2846df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2847df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2848d62866d3SStefano Zampini 2849df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2850d62866d3SStefano Zampini } 2851ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2852304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2853304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2854684f6988SStefano Zampini } 28556816873aSStefano Zampini /* free Neumann problem's matrix */ 28566816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2857304d26faSStefano Zampini 2858304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 28590fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2860684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2861684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 28621575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2863684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2864684f6988SStefano Zampini } 2865684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 28660fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 28670fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 28680fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 28690fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 28700fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2871304d26faSStefano Zampini /* need to be adapted? */ 2872b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2873b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2874b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2875304d26faSStefano Zampini /* print info */ 2876304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2877e604994aSStefano 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); 2878304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2879304d26faSStefano Zampini } 2880b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2881298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2882304d26faSStefano Zampini } 2883684f6988SStefano Zampini } 2884684f6988SStefano Zampini if (neumann) { /* Neumann */ 28856816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 28860fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 28870fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 28880fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 28890fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 28900fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2891304d26faSStefano Zampini /* need to be adapted? */ 2892b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2893b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2894304d26faSStefano Zampini /* print info */ 2895304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2896e604994aSStefano 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); 2897304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2898304d26faSStefano Zampini } 2899b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2900298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2901304d26faSStefano Zampini } 29020fccc4e9SStefano Zampini } 2903684f6988SStefano Zampini } 2904304d26faSStefano Zampini PetscFunctionReturn(0); 2905304d26faSStefano Zampini } 2906304d26faSStefano Zampini 2907304d26faSStefano Zampini #undef __FUNCT__ 2908ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 290980677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2910674ae819SStefano Zampini { 2911674ae819SStefano Zampini PetscErrorCode ierr; 2912674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2913be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2914674ae819SStefano Zampini 2915674ae819SStefano Zampini PetscFunctionBegin; 2916df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 291780677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 291820c7b377SStefano Zampini } 291980677318SStefano Zampini if (!pcbddc->switch_static) { 292080677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 292180677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 292280677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 292320c7b377SStefano Zampini } 2924df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 292580677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 292680677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 292720c7b377SStefano Zampini } else { 2928df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2929be83ff47SStefano Zampini 2930df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2931df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 293220c7b377SStefano Zampini } 2933be83ff47SStefano Zampini } else { 293480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 293580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 293680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 293780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 293880677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 293980677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 294080677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 294180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 294280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2943674ae819SStefano Zampini } 2944674ae819SStefano Zampini } 294572b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 294680677318SStefano Zampini if (applytranspose) { 294780677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 294880677318SStefano Zampini } else { 294980677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 295080677318SStefano Zampini } 2951be83ff47SStefano Zampini } else { 2952df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2953be83ff47SStefano Zampini 2954be83ff47SStefano Zampini if (applytranspose) { 2955df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2956be83ff47SStefano Zampini } else { 2957df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 2958be83ff47SStefano Zampini } 2959be83ff47SStefano Zampini } 296080677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 296180677318SStefano Zampini if (!pcbddc->switch_static) { 2962df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 296380677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 296480677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2965be83ff47SStefano Zampini } else { 2966df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2967be83ff47SStefano Zampini 2968df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2969df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2970be83ff47SStefano Zampini } 297180677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 297280677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 297380677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 297480677318SStefano Zampini } 297580677318SStefano Zampini } else { 297680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297880677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 297980677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298080677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 298180677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 298280677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 298380677318SStefano Zampini } 298480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298680677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 298780677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2988674ae819SStefano Zampini } 2989674ae819SStefano Zampini PetscFunctionReturn(0); 2990674ae819SStefano Zampini } 2991674ae819SStefano Zampini 2992dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 2993674ae819SStefano Zampini #undef __FUNCT__ 2994674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 2995dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 2996674ae819SStefano Zampini { 2997674ae819SStefano Zampini PetscErrorCode ierr; 2998674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2999674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3000674ae819SStefano Zampini const PetscScalar zero = 0.0; 3001674ae819SStefano Zampini 3002674ae819SStefano Zampini PetscFunctionBegin; 3003dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 30044fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3005dc359a40SStefano Zampini if (applytranspose) { 3006674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 30078eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3008dc359a40SStefano Zampini } else { 3009674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3010674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 301115aaf578SStefano Zampini } 30124fee134fSStefano Zampini } else { 30134fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 30144fee134fSStefano Zampini } 3015efc2fbd9SStefano Zampini 3016efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 30174f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3018efc2fbd9SStefano Zampini PetscScalar *array; 30194f1b2e48SStefano Zampini PetscInt j; 3020efc2fbd9SStefano Zampini 3021efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30224f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3023efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3024efc2fbd9SStefano Zampini } 3025efc2fbd9SStefano Zampini 302612edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 302712edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 302812edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 302912edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 303012edc857SStefano Zampini 30319f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 303212edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 303312edc857SStefano Zampini if (pcbddc->coarse_ksp) { 303451694757SStefano Zampini Mat coarse_mat; 3035964fefecSStefano Zampini Vec rhs,sol; 303651694757SStefano Zampini MatNullSpace nullsp; 3037964fefecSStefano Zampini 3038964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3039964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 304051694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 304151694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 304251694757SStefano Zampini if (nullsp) { 304351694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 304451694757SStefano Zampini } 304512edc857SStefano Zampini if (applytranspose) { 3046964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 304712edc857SStefano Zampini } else { 3048964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 304912edc857SStefano Zampini } 305051694757SStefano Zampini if (nullsp) { 305151694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 305251694757SStefano Zampini } 305312edc857SStefano Zampini } 3054674ae819SStefano Zampini 3055674ae819SStefano Zampini /* Local solution on R nodes */ 30564fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 305780677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 30589f00e9b4SStefano Zampini } 30599f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 30609f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 306112edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3062674ae819SStefano Zampini 30634fee134fSStefano Zampini /* Sum contributions from the two levels */ 30644fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3065dc359a40SStefano Zampini if (applytranspose) { 3066dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3067dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3068dc359a40SStefano Zampini } else { 3069674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 30708eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3071dc359a40SStefano Zampini } 3072efc2fbd9SStefano Zampini /* store p0 */ 30734f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3074efc2fbd9SStefano Zampini PetscScalar *array; 30754f1b2e48SStefano Zampini PetscInt j; 3076efc2fbd9SStefano Zampini 3077efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30784f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3079efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3080efc2fbd9SStefano Zampini } 30814fee134fSStefano Zampini } else { /* expand the coarse solution */ 30824fee134fSStefano Zampini if (applytranspose) { 30834fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 30844fee134fSStefano Zampini } else { 30854fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 30864fee134fSStefano Zampini } 30874fee134fSStefano Zampini } 3088674ae819SStefano Zampini PetscFunctionReturn(0); 3089674ae819SStefano Zampini } 3090674ae819SStefano Zampini 3091674ae819SStefano Zampini #undef __FUNCT__ 3092674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 309312edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3094674ae819SStefano Zampini { 3095674ae819SStefano Zampini PetscErrorCode ierr; 3096674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 309758da7f69SStefano Zampini PetscScalar *array; 309812edc857SStefano Zampini Vec from,to; 3099674ae819SStefano Zampini 3100674ae819SStefano Zampini PetscFunctionBegin; 310112edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 310212edc857SStefano Zampini from = pcbddc->coarse_vec; 310312edc857SStefano Zampini to = pcbddc->vec1_P; 310412edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 310512edc857SStefano Zampini Vec tvec; 310658da7f69SStefano Zampini 310758da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 310858da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 310912edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 311058da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 311158da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 311258da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 311312edc857SStefano Zampini } 311412edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 311512edc857SStefano Zampini from = pcbddc->vec1_P; 311612edc857SStefano Zampini to = pcbddc->coarse_vec; 311712edc857SStefano Zampini } 311812edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3119674ae819SStefano Zampini PetscFunctionReturn(0); 3120674ae819SStefano Zampini } 3121674ae819SStefano Zampini 3122674ae819SStefano Zampini #undef __FUNCT__ 3123674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 312412edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3125674ae819SStefano Zampini { 3126674ae819SStefano Zampini PetscErrorCode ierr; 3127674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 312858da7f69SStefano Zampini PetscScalar *array; 312912edc857SStefano Zampini Vec from,to; 3130674ae819SStefano Zampini 3131674ae819SStefano Zampini PetscFunctionBegin; 313212edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 313312edc857SStefano Zampini from = pcbddc->coarse_vec; 313412edc857SStefano Zampini to = pcbddc->vec1_P; 313512edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 313612edc857SStefano Zampini from = pcbddc->vec1_P; 313712edc857SStefano Zampini to = pcbddc->coarse_vec; 313812edc857SStefano Zampini } 313912edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 314012edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 314112edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 314212edc857SStefano Zampini Vec tvec; 314358da7f69SStefano Zampini 314412edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 314558da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 314658da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 314758da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 314858da7f69SStefano Zampini } 314958da7f69SStefano Zampini } else { 315058da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 315158da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 315212edc857SStefano Zampini } 315312edc857SStefano Zampini } 3154674ae819SStefano Zampini PetscFunctionReturn(0); 3155674ae819SStefano Zampini } 3156674ae819SStefano Zampini 3157984c4197SStefano Zampini /* uncomment for testing purposes */ 3158984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3159674ae819SStefano Zampini #undef __FUNCT__ 3160674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3161674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3162674ae819SStefano Zampini { 3163674ae819SStefano Zampini PetscErrorCode ierr; 3164674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3165674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3166674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3167984c4197SStefano Zampini /* one and zero */ 3168984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3169984c4197SStefano Zampini /* space to store constraints and their local indices */ 31709162d606SStefano Zampini PetscScalar *constraints_data; 31719162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 31729162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 31739162d606SStefano Zampini PetscInt *constraints_n; 3174984c4197SStefano Zampini /* iterators */ 3175b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3176984c4197SStefano Zampini /* BLAS integers */ 3177e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3178e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3179c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3180727cdba6SStefano Zampini /* reuse */ 31810e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 31820e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3183984c4197SStefano Zampini /* change of basis */ 3184b3d85658SStefano Zampini PetscBool qr_needed; 31859162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3186984c4197SStefano Zampini /* auxiliary stuff */ 318764efe560SStefano Zampini PetscInt *nnz,*is_indices; 31888a0068c3SStefano Zampini PetscInt ncc; 3189984c4197SStefano Zampini /* some quantities */ 319045a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3191a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3192984c4197SStefano Zampini 3193674ae819SStefano Zampini PetscFunctionBegin; 31948e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 31958e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 31968e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3197088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3198088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 31990e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 32000e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 32010e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32020e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32030e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3204088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3205cf5a6209SStefano Zampini 3206cf5a6209SStefano Zampini /* print some info */ 3207cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3208cf5a6209SStefano Zampini IS vertices; 3209cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3210cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3211cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3212cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 32131575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3214cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3215cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3216fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3217fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3218cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32191575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3220cf5a6209SStefano Zampini } 3221cf5a6209SStefano Zampini 3222cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 32239162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3224cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3225cf5a6209SStefano Zampini const Vec *nearnullvecs; 3226cf5a6209SStefano Zampini Vec *localnearnullsp; 3227cf5a6209SStefano Zampini PetscScalar *array; 3228cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3229cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3230674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3231b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3232674ae819SStefano Zampini PetscScalar *work; 3233674ae819SStefano Zampini PetscReal *singular_vals; 3234674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3235674ae819SStefano Zampini PetscReal *rwork; 3236674ae819SStefano Zampini #endif 3237674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3238674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3239674ae819SStefano Zampini #else 3240964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3241964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3242674ae819SStefano Zampini #endif 3243674ae819SStefano Zampini 3244674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3245d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3246d06fc5fdSStefano Zampini /* free unneeded index sets */ 3247d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3248d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3249674ae819SStefano Zampini } 3250d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3251d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3252d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3253d06fc5fdSStefano Zampini } 3254d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3255d06fc5fdSStefano Zampini n_ISForEdges = 0; 3256d06fc5fdSStefano Zampini } 3257d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3258d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3259d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3260d06fc5fdSStefano Zampini } 3261d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3262d06fc5fdSStefano Zampini n_ISForFaces = 0; 3263d06fc5fdSStefano Zampini } 326470022509SStefano Zampini 326570022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 326670022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 326770022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 326870022509SStefano Zampini if (pcbddc->NullSpace) { 326970022509SStefano Zampini PetscBool tbool[2],gbool[2]; 327070022509SStefano Zampini 327170022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3272b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3273d06fc5fdSStefano Zampini if (!ISForEdges) { 3274d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3275d06fc5fdSStefano Zampini } 3276b8ffe317SStefano Zampini } 3277d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3278d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3279d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3280d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3281d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 328298a51de6SStefano Zampini } 328370022509SStefano Zampini #endif 328408122e43SStefano Zampini 3285674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3286674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3287674ae819SStefano Zampini if (nearnullsp) { 3288674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3289f4ddd8eeSStefano Zampini /* remove any stored info */ 3290f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3291f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3292f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3293f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3294f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3295473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3296f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3297f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3298f4ddd8eeSStefano Zampini } 3299984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3300984c4197SStefano Zampini nnsp_size = 0; 3301674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3302674ae819SStefano Zampini } 3303984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3304984c4197SStefano Zampini max_constraints = nnsp_size; 3305984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3306984c4197SStefano Zampini 3307674ae819SStefano Zampini /* 3308674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 33099162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 33109162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 33119162d606SStefano Zampini There can be multiple constraints per connected component 3312674ae819SStefano Zampini */ 3313674ae819SStefano Zampini n_vertices = 0; 3314674ae819SStefano Zampini if (ISForVertices) { 3315674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3316674ae819SStefano Zampini } 33179162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 33189162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 33199162d606SStefano Zampini 33209162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 33219162d606SStefano Zampini total_counts *= max_constraints; 3322674ae819SStefano Zampini total_counts += n_vertices; 33234641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 33249162d606SStefano Zampini 3325674ae819SStefano Zampini total_counts = 0; 3326674ae819SStefano Zampini max_size_of_constraint = 0; 3327674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 33289162d606SStefano Zampini IS used_is; 3329674ae819SStefano Zampini if (i<n_ISForEdges) { 33309162d606SStefano Zampini used_is = ISForEdges[i]; 3331674ae819SStefano Zampini } else { 33329162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3333674ae819SStefano Zampini } 33349162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3335674ae819SStefano Zampini total_counts += j; 3336674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3337674ae819SStefano Zampini } 33389162d606SStefano 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); 33399162d606SStefano Zampini 3340984c4197SStefano Zampini /* get local part of global near null space vectors */ 3341785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3342984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3343984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3344e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3345e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3346984c4197SStefano Zampini } 3347674ae819SStefano Zampini 3348242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3349242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3350a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3351242a89d7SStefano Zampini 3352984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3353a773dcb8SStefano Zampini if (!skip_lapack) { 3354674ae819SStefano Zampini PetscScalar temp_work; 3355911cabfeSStefano Zampini 3356674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3357984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3358785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3359785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3360785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3361674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3362785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3363674ae819SStefano Zampini #endif 3364674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3365c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3366c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3367674ae819SStefano Zampini lwork = -1; 3368674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3369674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3370c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3371674ae819SStefano Zampini #else 3372c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3373674ae819SStefano Zampini #endif 3374674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3375984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3376674ae819SStefano Zampini #else /* on missing GESVD */ 3377674ae819SStefano Zampini /* SVD */ 3378674ae819SStefano Zampini PetscInt max_n,min_n; 3379674ae819SStefano Zampini max_n = max_size_of_constraint; 3380984c4197SStefano Zampini min_n = max_constraints; 3381984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3382674ae819SStefano Zampini min_n = max_size_of_constraint; 3383984c4197SStefano Zampini max_n = max_constraints; 3384674ae819SStefano Zampini } 3385785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3386674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3387785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3388674ae819SStefano Zampini #endif 3389674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3390674ae819SStefano Zampini lwork = -1; 3391e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3392e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3393b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3394674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3395674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 33969162d606SStefano 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)); 3397674ae819SStefano Zampini #else 33989162d606SStefano 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)); 3399674ae819SStefano Zampini #endif 3400674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3401984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3402984c4197SStefano Zampini #endif /* on missing GESVD */ 3403674ae819SStefano Zampini /* Allocate optimal workspace */ 3404674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3405854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3406674ae819SStefano Zampini } 3407674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3408674ae819SStefano Zampini total_counts = 0; 34099162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 34109162d606SStefano Zampini constraints_data_ptr[0] = 0; 3411674ae819SStefano Zampini /* vertices */ 34129162d606SStefano Zampini if (n_vertices) { 3413674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 34149162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3415674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 34169162d606SStefano Zampini constraints_n[total_counts] = 1; 34179162d606SStefano Zampini constraints_data[total_counts] = 1.0; 34189162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 34199162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3420674ae819SStefano Zampini total_counts++; 3421674ae819SStefano Zampini } 3422674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3423674ae819SStefano Zampini n_vertices = total_counts; 3424674ae819SStefano Zampini } 3425984c4197SStefano Zampini 3426674ae819SStefano Zampini /* edges and faces */ 34279162d606SStefano Zampini total_counts_cc = total_counts; 3428911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 34299162d606SStefano Zampini IS used_is; 34309162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 34319162d606SStefano Zampini 3432911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 34339162d606SStefano Zampini used_is = ISForEdges[ncc]; 3434984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3435674ae819SStefano Zampini } else { 34369162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3437984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3438674ae819SStefano Zampini } 3439674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 34409162d606SStefano Zampini 34419162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 34429162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3443984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3444984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3445674ae819SStefano Zampini if (nnsp_has_cnst) { 34465b08dc53SStefano Zampini PetscScalar quad_value; 34479162d606SStefano Zampini 34489162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34499162d606SStefano Zampini idxs_copied = PETSC_TRUE; 34509162d606SStefano Zampini 3451a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3452674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3453a773dcb8SStefano Zampini } else { 3454a773dcb8SStefano Zampini quad_value = 1.0; 3455a773dcb8SStefano Zampini } 3456674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34579162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3458674ae819SStefano Zampini } 34599162d606SStefano Zampini temp_constraints++; 3460674ae819SStefano Zampini total_counts++; 3461674ae819SStefano Zampini } 3462674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3463984c4197SStefano Zampini PetscReal real_value; 34649162d606SStefano Zampini PetscScalar *ptr_to_data; 34659162d606SStefano Zampini 3466984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 34679162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3468674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 34699162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3470674ae819SStefano Zampini } 3471984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3472984c4197SStefano Zampini /* check if array is null on the connected component */ 3473e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34749162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 34755b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3476674ae819SStefano Zampini temp_constraints++; 3477674ae819SStefano Zampini total_counts++; 34789162d606SStefano Zampini if (!idxs_copied) { 34799162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 34809162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3481674ae819SStefano Zampini } 3482674ae819SStefano Zampini } 34839162d606SStefano Zampini } 34849162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 348545a1bb75SStefano Zampini valid_constraints = temp_constraints; 3486eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3487a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 34889162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 34899162d606SStefano Zampini 34909162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3491a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 34929162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3493a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 34949162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3495a773dcb8SStefano Zampini } else { /* perform SVD */ 3496984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 34979162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3498674ae819SStefano Zampini 3499674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3500984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3501984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3502984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3503984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3504984c4197SStefano Zampini from that computed using LAPACKgesvd 3505984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3506984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3507984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3508674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3509e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3510984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3511674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3512674ae819SStefano Zampini for (k=0;k<j+1;k++) { 35139162d606SStefano 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)); 3514674ae819SStefano Zampini } 3515674ae819SStefano Zampini } 3516e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3517e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3518e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3519674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3520c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3521674ae819SStefano Zampini #else 3522c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3523674ae819SStefano Zampini #endif 3524674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3525984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3526984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3527674ae819SStefano Zampini j = 0; 3528984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3529674ae819SStefano Zampini total_counts = total_counts-j; 353045a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3531e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3532c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3533c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3534c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3535c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3536c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3537c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3538674ae819SStefano Zampini if (j<temp_constraints) { 3539984c4197SStefano Zampini PetscInt ii; 3540984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3541674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 35429162d606SStefano 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)); 3543674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3544984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3545674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 35469162d606SStefano 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]; 3547674ae819SStefano Zampini } 3548674ae819SStefano Zampini } 3549674ae819SStefano Zampini } 3550674ae819SStefano Zampini #else /* on missing GESVD */ 3551e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3552e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3553b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3554674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3555674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 35569162d606SStefano 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)); 3557674ae819SStefano Zampini #else 35589162d606SStefano 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)); 3559674ae819SStefano Zampini #endif 3560984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3561674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3562984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3563e310c8b4SStefano Zampini k = temp_constraints; 3564e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3565674ae819SStefano Zampini j = 0; 3566e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 356745a1bb75SStefano Zampini valid_constraints = k-j; 3568911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3569984c4197SStefano Zampini #endif /* on missing GESVD */ 3570674ae819SStefano Zampini } 3571a773dcb8SStefano Zampini } 35729162d606SStefano Zampini /* update pointers information */ 35739162d606SStefano Zampini if (valid_constraints) { 35749162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 35759162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 35769162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 35779162d606SStefano Zampini /* set change_of_basis flag */ 357845a1bb75SStefano Zampini if (boolforchange) { 3579b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 35809162d606SStefano Zampini } 3581b3d85658SStefano Zampini total_counts_cc++; 358245a1bb75SStefano Zampini } 358345a1bb75SStefano Zampini } 3584984c4197SStefano Zampini /* free workspace */ 35858f1c130eSStefano Zampini if (!skip_lapack) { 3586984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3587984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3588984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3589984c4197SStefano Zampini #endif 3590984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3591984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3592984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3593984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3594984c4197SStefano Zampini #endif 3595984c4197SStefano Zampini } 3596984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3597984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3598984c4197SStefano Zampini } 3599984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3600cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3601cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3602cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3603cf5a6209SStefano Zampini } 3604cf5a6209SStefano Zampini if (n_ISForFaces) { 3605cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3606cf5a6209SStefano Zampini } 3607cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3608cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3609cf5a6209SStefano Zampini } 3610cf5a6209SStefano Zampini if (n_ISForEdges) { 3611cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3612cf5a6209SStefano Zampini } 3613cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 361408122e43SStefano Zampini } else { 361508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3616984c4197SStefano Zampini 361708122e43SStefano Zampini total_counts = 0; 361808122e43SStefano Zampini n_vertices = 0; 3619d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3620d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 362108122e43SStefano Zampini } 362208122e43SStefano Zampini max_constraints = 0; 36239162d606SStefano Zampini total_counts_cc = 0; 362408122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 362508122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 36269162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 362708122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 362808122e43SStefano Zampini } 36299162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 36309162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 36319162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 36329162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 363374d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 36349162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 36359162d606SStefano Zampini total_counts_cc = 0; 36369162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 36379162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 36389162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 363908122e43SStefano Zampini } 364008122e43SStefano Zampini } 36419162d606SStefano Zampini #if 0 36429162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 36439162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36449162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 36459162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 36469162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 36479162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 36489162d606SStefano Zampini } 36499162d606SStefano Zampini printf("\n"); 36509162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 36519162d606SStefano Zampini } 36521b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 36538bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 36541b968477SStefano Zampini } 36551b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 36568bec7fa6SStefano 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]); 36571b968477SStefano Zampini } 365808122e43SStefano Zampini #endif 365908122e43SStefano Zampini 36608bec7fa6SStefano Zampini max_size_of_constraint = 0; 36619162d606SStefano 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]); 36629162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 366308122e43SStefano Zampini /* Change of basis */ 3664b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 366508122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 366608122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 366708122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3668b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 366908122e43SStefano Zampini } 367008122e43SStefano Zampini } 367108122e43SStefano Zampini } 367208122e43SStefano Zampini } 3673984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 36744f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 367508122e43SStefano Zampini 36769162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 36779162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 36789162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 36799162d606SStefano 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); 368008122e43SStefano Zampini } 3681674ae819SStefano Zampini 3682674ae819SStefano Zampini /* Create constraint matrix */ 3683674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 368416f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3685984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3686984c4197SStefano Zampini 3687984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3688a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3689a717540cSStefano Zampini qr_needed = PETSC_FALSE; 369074d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3691984c4197SStefano Zampini total_primal_vertices=0; 3692b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 36939162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 36949162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 369572b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 36969162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3697b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 369864efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 36999162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37009162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3701a717540cSStefano Zampini } 3702b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 370374d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3704a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3705a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3706a717540cSStefano Zampini } 3707fa434743SStefano Zampini } else { 3708b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3709fa434743SStefano Zampini } 3710a717540cSStefano Zampini } 3711b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3712b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3713674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 371470022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3715b3d85658SStefano Zampini 37164f1b2e48SStefano 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); 37170e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 37180e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3719984c4197SStefano Zampini 3720984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 372174d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3722785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3723984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 372474d5cdf7SStefano Zampini 3725984c4197SStefano Zampini j = total_primal_vertices; 372674d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3727b3d85658SStefano Zampini cum = total_primal_vertices; 37289162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37294641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3730b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3731b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3732b3d85658SStefano Zampini cum++; 37339162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 373474d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 373574d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 373674d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 373774d5cdf7SStefano Zampini } 37389162d606SStefano Zampini j += constraints_n[i]; 3739674ae819SStefano Zampini } 3740674ae819SStefano Zampini } 3741674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3742674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3743088faed8SStefano Zampini 3744674ae819SStefano Zampini /* set values in constraint matrix */ 3745984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 37460e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3747674ae819SStefano Zampini } 3748984c4197SStefano Zampini total_counts = total_primal_vertices; 37499162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 37504641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 37519162d606SStefano Zampini PetscInt *cols; 37529162d606SStefano Zampini 37539162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 37549162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 37559162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37569162d606SStefano Zampini PetscInt row = total_counts+k; 37579162d606SStefano Zampini PetscScalar *vals; 37589162d606SStefano Zampini 37599162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 37609162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 37619162d606SStefano Zampini } 37629162d606SStefano Zampini total_counts += constraints_n[i]; 3763674ae819SStefano Zampini } 3764674ae819SStefano Zampini } 3765674ae819SStefano Zampini /* assembling */ 3766674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3767674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3768088faed8SStefano Zampini 3769984c4197SStefano Zampini /* 377045a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3771984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3772984c4197SStefano Zampini */ 3773674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3774674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3775026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3776984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3777984c4197SStefano Zampini /* working stuff for GEQRF */ 377881d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3779984c4197SStefano Zampini PetscBLASInt lqr_work; 3780984c4197SStefano Zampini /* working stuff for UNGQR */ 3781984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3782984c4197SStefano Zampini PetscBLASInt lgqr_work; 3783984c4197SStefano Zampini /* working stuff for TRTRS */ 3784984c4197SStefano Zampini PetscScalar *trs_rhs; 37853f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3786984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3787984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3788984c4197SStefano Zampini PetscScalar *start_vals; 3789984c4197SStefano Zampini /* working stuff for values insertion */ 37904641a718SStefano Zampini PetscBT is_primal; 379164efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3792906d46d4SStefano Zampini /* matrix sizes */ 3793906d46d4SStefano Zampini PetscInt global_size,local_size; 3794906d46d4SStefano Zampini /* temporary change of basis */ 3795906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3796cf5a6209SStefano Zampini /* extra space for debugging */ 3797cf5a6209SStefano Zampini PetscScalar *dbg_work; 3798984c4197SStefano Zampini 3799906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3800906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 380116f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3802bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3803906d46d4SStefano Zampini /* nonzeros for local mat */ 3804bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3805bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 38069162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3807a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 38089162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3809a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 38109162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3811a717540cSStefano Zampini } else { 38129162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 38139162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3814a717540cSStefano Zampini } 3815a717540cSStefano Zampini } 3816a717540cSStefano Zampini } 3817906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3818bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3819a717540cSStefano Zampini /* Set initial identity in the matrix */ 3820bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3821906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3822a717540cSStefano Zampini } 3823a717540cSStefano Zampini 3824a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3825a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3826a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3827a717540cSStefano Zampini } 3828a717540cSStefano Zampini 3829a717540cSStefano Zampini 3830a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3831a717540cSStefano Zampini /* 3832a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3833a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3834a717540cSStefano Zampini 3835a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3836a717540cSStefano Zampini 3837a6b551f4SStefano 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) 3838a6b551f4SStefano Zampini 3839a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3840a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3841a717540cSStefano Zampini | ... | 3842a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3843a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3844a717540cSStefano Zampini 3845a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3846a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3847a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3848a6b551f4SStefano Zampini 3849a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3850a717540cSStefano Zampini */ 3851a717540cSStefano Zampini if (qr_needed) { 3852984c4197SStefano Zampini /* space to store Q */ 3853854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3854984c4197SStefano Zampini /* first we issue queries for optimal work */ 38553f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38563f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 38573f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3858984c4197SStefano Zampini lqr_work = -1; 38593f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3860984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3861984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3862785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3863984c4197SStefano Zampini lgqr_work = -1; 38643f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 38653f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 38663f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 38673f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 38683f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 38693f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3870984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3871984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3872785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3873984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3874785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3875984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3876785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3877a717540cSStefano Zampini /* allocating workspace for check */ 3878a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3879cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3880a717540cSStefano Zampini } 3881a717540cSStefano Zampini } 3882984c4197SStefano Zampini /* array to store whether a node is primal or not */ 38834641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3884473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 38850e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 388639e2fb2aSStefano Zampini if (i != total_primal_vertices) { 388739e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 38884641a718SStefano Zampini } 388939e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 389039e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 389139e2fb2aSStefano Zampini } 389239e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3893984c4197SStefano Zampini 3894a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 38959162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 38969162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 38974641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3898984c4197SStefano Zampini /* get constraint info */ 38999162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3900984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3901984c4197SStefano Zampini 3902984c4197SStefano Zampini if (pcbddc->dbg_flag) { 39039162d606SStefano 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); 3904674ae819SStefano Zampini } 3905984c4197SStefano Zampini 3906fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3907a717540cSStefano Zampini 3908a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3909a717540cSStefano Zampini if (pcbddc->dbg_flag) { 39109162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3911a717540cSStefano Zampini } 3912984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 39139162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3914984c4197SStefano Zampini 3915984c4197SStefano Zampini /* compute QR decomposition of constraints */ 39163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39173f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39183f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3919674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39203f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3921984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3922674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3923984c4197SStefano Zampini 3924984c4197SStefano Zampini /* explictly compute R^-T */ 3925984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 3926984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 39273f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39283f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 39293f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39303f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 3931984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39323f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 3933984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 3934984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3935984c4197SStefano Zampini 3936a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 39373f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39383f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 39393f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39403f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3941984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39423f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 3943984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 3944984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3945984c4197SStefano Zampini 3946984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 3947984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 3948984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 39493f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39503f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39513f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 39523f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39533f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 39543f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3955984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39569162d606SStefano 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)); 3957984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 39589162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3959984c4197SStefano Zampini 3960984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 39619162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 3962984c4197SStefano Zampini /* insert cols for primal dofs */ 3963984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 3964984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 39659162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 3966906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3967984c4197SStefano Zampini } 3968984c4197SStefano Zampini /* insert cols for dual dofs */ 3969984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 39709162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 3971984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 39729162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 3973906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 3974984c4197SStefano Zampini j++; 3975674ae819SStefano Zampini } 3976674ae819SStefano Zampini } 3977984c4197SStefano Zampini 3978984c4197SStefano Zampini /* check change of basis */ 3979984c4197SStefano Zampini if (pcbddc->dbg_flag) { 3980984c4197SStefano Zampini PetscInt ii,jj; 3981984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 3982c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 3983c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3984c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 3985c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3986c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 3987c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 3988984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3989cf5a6209SStefano 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)); 3990984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3991984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 3992984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 3993cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 3994cf5a6209SStefano 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; 3995674ae819SStefano Zampini } 3996674ae819SStefano Zampini } 3997984c4197SStefano Zampini if (!valid_qr) { 399822d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 3999984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4000984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4001cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4002cf5a6209SStefano 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])); 4003674ae819SStefano Zampini } 4004cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4005cf5a6209SStefano 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])); 4006984c4197SStefano Zampini } 4007984c4197SStefano Zampini } 4008984c4197SStefano Zampini } 4009674ae819SStefano Zampini } else { 401022d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4011674ae819SStefano Zampini } 4012674ae819SStefano Zampini } 4013a717540cSStefano Zampini } else { /* simple transformation block */ 4014a717540cSStefano Zampini PetscInt row,col; 4015a6b551f4SStefano Zampini PetscScalar val,norm; 4016a6b551f4SStefano Zampini 4017a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40189162d606SStefano 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)); 4019a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 40209162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 40219162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4022bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 40239162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4024906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 40259162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4026a717540cSStefano Zampini } else { 4027a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 40289162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4029a717540cSStefano Zampini if (row != col) { 40309162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4031a717540cSStefano Zampini } else { 40329162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4033a717540cSStefano Zampini } 4034906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4035a717540cSStefano Zampini } 4036a717540cSStefano Zampini } 4037a717540cSStefano Zampini } 403898a51de6SStefano Zampini if (pcbddc->dbg_flag) { 403922d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4040a717540cSStefano Zampini } 4041674ae819SStefano Zampini } 4042984c4197SStefano Zampini } else { 4043984c4197SStefano Zampini if (pcbddc->dbg_flag) { 40449162d606SStefano 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); 4045674ae819SStefano Zampini } 4046674ae819SStefano Zampini } 4047674ae819SStefano Zampini } 4048a717540cSStefano Zampini 4049a717540cSStefano Zampini /* free workspace */ 4050a717540cSStefano Zampini if (qr_needed) { 4051984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4052cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4053984c4197SStefano Zampini } 4054984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4055984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4056984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4057984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4058984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4059674ae819SStefano Zampini } 4060a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4061906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4062906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4063906d46d4SStefano Zampini 4064906d46d4SStefano Zampini /* assembling of global change of variable */ 406588c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4066bbb9e6c6SStefano Zampini Mat tmat; 406716f15bc4SStefano Zampini PetscInt bs; 406816f15bc4SStefano Zampini 4069906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4070906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4071bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4072bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4073bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4074bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 407516f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 407616f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4077906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4078bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4079bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4080bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4081bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4082bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4083e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4084e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4085bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4086bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 408788c03ad3SStefano Zampini 4088906d46d4SStefano Zampini /* check */ 4089906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4090906d46d4SStefano Zampini PetscReal error; 4091906d46d4SStefano Zampini Vec x,x_change; 4092906d46d4SStefano Zampini 4093906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4094906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4095906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4096906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4097e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4098e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4099bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4100e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4101e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4102906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4103906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4104906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4105906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4106bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4107906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4108906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4109906d46d4SStefano Zampini } 4110b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4111b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4112b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4113bf3a8328SStefano Zampini 4114bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4115bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4116bf3a8328SStefano Zampini } 4117b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4118ac632422SStefano Zampini Mat S_new,tmat; 4119bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4120bbb9e6c6SStefano Zampini 4121bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 41226816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4123bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4124bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4125bf3a8328SStefano Zampini IS is_V; 4126b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4127b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4128b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4129b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4130b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4131bf3a8328SStefano Zampini } 4132bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4133ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4134b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4135ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4136bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4137bf3a8328SStefano Zampini const PetscScalar *array; 4138bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4139bf3a8328SStefano Zampini PetscInt i,n_V; 4140bf3a8328SStefano Zampini 4141b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4142b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4143b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4144b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4145b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4146b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4147b087196eSStefano Zampini PetscScalar val; 4148b087196eSStefano Zampini PetscInt idx; 4149b087196eSStefano Zampini 4150b087196eSStefano Zampini idx = idxs_V[i]; 4151b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4152b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4153b087196eSStefano Zampini } 4154b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4155b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4156bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4157bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4158bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4159bf3a8328SStefano Zampini } 4160ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4161ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4162ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4163ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4164b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4165ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4166bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4167b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4168bf3a8328SStefano Zampini } 4169ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4170ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4171ac632422SStefano Zampini } 4172b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 417388c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4174b96c3477SStefano Zampini } 4175b96c3477SStefano Zampini } 4176906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 417772b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 417872b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 417972b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 418072b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 418172b8c272SStefano Zampini } 4182906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4183b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4184b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4185b9b85e73SStefano Zampini } 4186906d46d4SStefano Zampini 4187906d46d4SStefano Zampini /* set up change of basis context */ 4188906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4189906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4190906d46d4SStefano Zampini 4191906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4192906d46d4SStefano Zampini PetscInt global_size,local_size; 4193906d46d4SStefano Zampini 4194906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4195906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4196906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4197906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4198906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4199906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4200906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4201906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4202906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4203906d46d4SStefano Zampini } else { 4204906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4205906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4206906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4207906d46d4SStefano Zampini } 4208906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4209906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4210906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4211906d46d4SStefano Zampini } else { 4212906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4213906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4214906d46d4SStefano Zampini } 4215906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4216906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4217906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4218906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4219a8661815SStefano Zampini } else { 4220a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4221b9b85e73SStefano Zampini } 4222a717540cSStefano Zampini 422372b8c272SStefano Zampini if (!pcbddc->fake_change) { 42244f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 42254f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 42264f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 42274f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4228019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4229019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4230019a44ceSStefano Zampini pcbddc->local_primal_size++; 4231019a44ceSStefano Zampini } 4232019a44ceSStefano Zampini 4233019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4234727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4235727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 42369f47a83aSStefano 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); 4237c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 42380e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 42399f47a83aSStefano 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); 4240727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4241727cdba6SStefano Zampini } 42420e6343abSStefano Zampini } 4243727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4244727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 424572b8c272SStefano Zampini } 424672b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4247727cdba6SStefano Zampini 4248a717540cSStefano Zampini /* flush dbg viewer */ 4249b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4250b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4251b8ffe317SStefano Zampini } 4252a717540cSStefano Zampini 4253e310c8b4SStefano Zampini /* free workspace */ 4254a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 42554641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 425608122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 42579162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 42589162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 425908122e43SStefano Zampini } else { 42609162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 42619162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 42629162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 426308122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 426408122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 42659162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 42669162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 426708122e43SStefano Zampini } 4268674ae819SStefano Zampini PetscFunctionReturn(0); 4269674ae819SStefano Zampini } 4270674ae819SStefano Zampini 4271674ae819SStefano Zampini #undef __FUNCT__ 4272674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4273674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4274674ae819SStefano Zampini { 4275674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4276674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4277674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 4278*14f95afaSStefano Zampini PetscInt ierr,i,N; 4279674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4280674ae819SStefano Zampini 4281674ae819SStefano Zampini PetscFunctionBegin; 42828e61c736SStefano Zampini /* Reset previously computed graph */ 42838e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4284674ae819SStefano Zampini /* Init local Graph struct */ 42857fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 42863bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4287674ae819SStefano Zampini 4288575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 42895099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 42905099eff2SStefano 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); 4291575ad6abSStefano Zampini } 42929577ea80SStefano Zampini 4293674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4294d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 42954d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 42964d379d7bSStefano Zampini PetscInt nvtxs; 4297e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4298674ae819SStefano Zampini 42992fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 43002fffb893SStefano Zampini if (flg_row) { 43014d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4302b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 43032fffb893SStefano Zampini } 43042fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4305674ae819SStefano Zampini } 43069b28b941SStefano Zampini if (pcbddc->dbg_flag) { 43079b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4308674ae819SStefano Zampini } 4309674ae819SStefano Zampini 431063602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 431163602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 431263602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 431395ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 431463602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4315e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 431663602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4317674ae819SStefano Zampini } 431863602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 431963602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 432063602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4321674ae819SStefano Zampini } 432263602bcaSStefano Zampini } else { 432363602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 432463602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4325854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 432663602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 432763602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 432863602bcaSStefano Zampini } 432963602bcaSStefano Zampini } 4330674ae819SStefano Zampini } 4331674ae819SStefano Zampini 4332674ae819SStefano Zampini /* Setup of Graph */ 4333785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4334e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4335785d1243SStefano Zampini } 4336785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4337e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4338785d1243SStefano Zampini } 433930368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 434030368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 434130368db7SStefano Zampini } 4342*14f95afaSStefano Zampini ierr = PCBDDCGraphSetUp(pcbddc->mat_graph,pcbddc->vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 4343674ae819SStefano Zampini 43444f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 43454f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 43464f1b2e48SStefano Zampini PetscInt *local_subs; 43474f1b2e48SStefano Zampini 43484f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 43494f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 43504f1b2e48SStefano Zampini const PetscInt *idxs; 43514f1b2e48SStefano Zampini PetscInt nl,j; 43524f1b2e48SStefano Zampini 43534f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 43544f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43554f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 43564f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 43574f1b2e48SStefano Zampini } 43584f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 43594f1b2e48SStefano Zampini } 43604f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 43614f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 43624f1b2e48SStefano Zampini } 43634f1b2e48SStefano Zampini 4364674ae819SStefano Zampini /* Graph's connected components analysis */ 4365674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4366674ae819SStefano Zampini 4367674ae819SStefano Zampini /* print some info to stdout */ 4368674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4369302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4370674ae819SStefano Zampini } 4371674ae819SStefano Zampini PetscFunctionReturn(0); 4372674ae819SStefano Zampini } 4373674ae819SStefano Zampini 4374dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4375674ae819SStefano Zampini #undef __FUNCT__ 4376674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4377dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4378674ae819SStefano Zampini { 4379dc456d91SStefano Zampini PetscSF sf; 4380dc456d91SStefano Zampini PetscLayout map; 4381dc456d91SStefano Zampini const PetscInt *idxs; 4382dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4383dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4384dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4385dc456d91SStefano Zampini PetscMPIInt commsize; 4386674ae819SStefano Zampini PetscBool first_found; 4387674ae819SStefano Zampini PetscErrorCode ierr; 4388674ae819SStefano Zampini 4389674ae819SStefano Zampini PetscFunctionBegin; 4390dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4391dc456d91SStefano Zampini if (subset_mult) { 4392dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4393dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4394dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4395674ae819SStefano Zampini } 4396dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4397dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4398dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4399dc456d91SStefano Zampini for (i=0;i<n;i++) { 4400dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4401dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4402674ae819SStefano Zampini } 4403dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4404dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4405dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4406dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4407dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4408dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4409dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4410dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4411dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4412dc456d91SStefano Zampini 4413dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4414dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4415dc456d91SStefano Zampini if (subset_mult) { 4416dc456d91SStefano Zampini const PetscInt* idxs_mult; 4417dc456d91SStefano Zampini 4418dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4419dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4420dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4421674ae819SStefano Zampini } else { 4422dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4423674ae819SStefano Zampini } 4424dc456d91SStefano Zampini /* local size of new subset */ 4425dc456d91SStefano Zampini n_n = 0; 4426dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4427dc456d91SStefano Zampini 4428dc456d91SStefano Zampini /* global indexes in layout */ 4429dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4430dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4431dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4432dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4433dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4434dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4435dc456d91SStefano Zampini 4436dc456d91SStefano Zampini /* reduce from leaves to roots */ 4437dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 443864a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 443964a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4440dc456d91SStefano Zampini 4441dc456d91SStefano Zampini /* count indexes in local part of layout */ 4442674ae819SStefano Zampini nlocals = 0; 4443674ae819SStefano Zampini first_index = -1; 4444674ae819SStefano Zampini first_found = PETSC_FALSE; 4445dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4446dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4447674ae819SStefano Zampini first_found = PETSC_TRUE; 4448674ae819SStefano Zampini first_index = i; 4449674ae819SStefano Zampini } 4450dc456d91SStefano Zampini nlocals += root_data[i]; 4451674ae819SStefano Zampini } 4452dc456d91SStefano Zampini 4453dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 44545fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4455dc456d91SStefano Zampini start = 0; 445664a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44575fa240b1SStefano Zampini #else 445864a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 44595fa240b1SStefano Zampini start = start-nlocals; 44605fa240b1SStefano Zampini #endif 44615fa240b1SStefano Zampini 4462dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4463dc456d91SStefano Zampini *N_n = start + nlocals; 4464dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4465dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4466674ae819SStefano Zampini } 44675fa240b1SStefano Zampini 44685fa240b1SStefano Zampini /* adapt root data with cumulative */ 4469674ae819SStefano Zampini if (first_found) { 4470dc456d91SStefano Zampini PetscInt old_index; 4471dc456d91SStefano Zampini 4472dc456d91SStefano Zampini root_data[first_index] += start; 4473674ae819SStefano Zampini old_index = first_index; 4474dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4475dc456d91SStefano Zampini if (root_data[i]) { 4476dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4477674ae819SStefano Zampini old_index = i; 4478674ae819SStefano Zampini } 4479674ae819SStefano Zampini } 4480674ae819SStefano Zampini } 4481dc456d91SStefano Zampini 4482dc456d91SStefano Zampini /* from roots to leaves */ 4483dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4484dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4485dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4486dc456d91SStefano Zampini 4487dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4488dc456d91SStefano Zampini if (subset_mult) { 4489dc456d91SStefano Zampini const PetscInt* idxs_mult; 4490dc456d91SStefano Zampini PetscInt cum; 4491dc456d91SStefano Zampini 4492dc456d91SStefano Zampini cum = 0; 4493dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4494dc456d91SStefano Zampini for (i=0;i<n;i++) { 4495dc456d91SStefano Zampini PetscInt j; 4496dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4497674ae819SStefano Zampini } 4498dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4499674ae819SStefano Zampini } else { 4500dc456d91SStefano Zampini for (i=0;i<n;i++) { 4501dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4502674ae819SStefano Zampini } 4503674ae819SStefano Zampini } 4504dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4505dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4506674ae819SStefano Zampini PetscFunctionReturn(0); 4507674ae819SStefano Zampini } 45089a7d3425SStefano Zampini 45099a7d3425SStefano Zampini #undef __FUNCT__ 45109a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 45119a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 45129a7d3425SStefano Zampini { 45139a7d3425SStefano Zampini PetscInt i,j; 45149a7d3425SStefano Zampini PetscScalar *alphas; 45159a7d3425SStefano Zampini PetscErrorCode ierr; 45169a7d3425SStefano Zampini 45179a7d3425SStefano Zampini PetscFunctionBegin; 45189a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4519785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 45209a7d3425SStefano Zampini for (i=0;i<n;i++) { 45219a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 45229a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 45239a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 45249a7d3425SStefano Zampini } 45259a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 45269a7d3425SStefano Zampini PetscFunctionReturn(0); 45279a7d3425SStefano Zampini } 45289a7d3425SStefano Zampini 4529e7931f94SStefano Zampini #undef __FUNCT__ 453070cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4531b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4532e7931f94SStefano Zampini { 453352e5ac9dSStefano Zampini IS ranks_send_to; 4534e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4535e7931f94SStefano Zampini PetscMPIInt size,rank,color; 453652e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 453752e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 45383837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 45392b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4540e7931f94SStefano Zampini PetscSubcomm subcomm; 454152e5ac9dSStefano Zampini PetscErrorCode ierr; 4542a57a6d2fSStefano Zampini 4543e7931f94SStefano Zampini PetscFunctionBegin; 45442b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 45452b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 45462b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4547e7931f94SStefano Zampini 4548e7931f94SStefano Zampini /* Get info on mapping */ 45493bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 45503bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4551e7931f94SStefano Zampini 4552e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4553785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4554e7931f94SStefano Zampini xadj[0] = 0; 4555e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4556785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4557785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4558e7931f94SStefano Zampini 45592b510759SStefano Zampini if (threshold) { 4560d023bfaeSStefano Zampini PetscInt xadj_count = 0; 45612b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4562d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4563d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4564d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4565d023bfaeSStefano Zampini xadj_count++; 4566e7931f94SStefano Zampini } 4567e7931f94SStefano Zampini } 4568d023bfaeSStefano Zampini xadj[1] = xadj_count; 4569c8587f34SStefano Zampini } else { 4570e7931f94SStefano Zampini if (xadj[1]) { 4571e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4572e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4573c8587f34SStefano Zampini } 4574e7931f94SStefano Zampini } 45753bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4576e7931f94SStefano Zampini if (use_square) { 4577e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4578e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4579e7931f94SStefano Zampini } 4580e7931f94SStefano Zampini } 4581e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4582e7931f94SStefano Zampini 45833837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4584e7931f94SStefano Zampini 4585e7931f94SStefano Zampini /* 4586e7931f94SStefano Zampini Restrict work on active processes only. 4587e7931f94SStefano Zampini */ 4588e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4589e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4590e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 45912b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4592d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4593e7931f94SStefano Zampini if (color) { 4594e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4595e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4596e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4597c8587f34SStefano Zampini } else { 459852e5ac9dSStefano Zampini Mat subdomain_adj; 459952e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 460052e5ac9dSStefano Zampini MatPartitioning partitioner; 460152e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 460252e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4603b0c7d250SStefano Zampini PetscBool aggregate; 4604b0c7d250SStefano Zampini 4605306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4606785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4607e7931f94SStefano Zampini prank = rank; 4608306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 46098002ef2cSStefano Zampini /* 4610e7931f94SStefano Zampini for (i=0;i<size;i++) { 4611e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4612c8587f34SStefano Zampini } 46138002ef2cSStefano Zampini */ 4614e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4615e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4616c8587f34SStefano Zampini } 4617e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4618b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4619b0c7d250SStefano Zampini if (aggregate) { 4620b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4621b0c7d250SStefano Zampini PetscMPIInt nrank; 4622b0c7d250SStefano Zampini PetscScalar *vals; 4623b0c7d250SStefano Zampini 4624b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4625b0c7d250SStefano Zampini lrows = 0; 4626b0c7d250SStefano Zampini if (nrank<redprocs) { 4627b0c7d250SStefano Zampini lrows = size/redprocs; 4628b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4629b0c7d250SStefano Zampini } 46305fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4631b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4632b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4633b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4634b0c7d250SStefano Zampini row = nrank; 4635b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4636b0c7d250SStefano Zampini cols = adjncy; 4637b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4638b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4639b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4640b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4641b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 464252e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 464352e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 464452e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4645b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4646b0c7d250SStefano Zampini } else { 4647306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4648b0c7d250SStefano Zampini } 464922b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4650e7931f94SStefano Zampini 4651e7931f94SStefano Zampini /* Partition */ 4652306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4653e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4654e7931f94SStefano Zampini if (use_vwgt) { 46553837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4656e7931f94SStefano Zampini v_wgt[0] = local_size; 4657e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4658c8587f34SStefano Zampini } 465928143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 466028143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4661e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4662e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 466322b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4664e7931f94SStefano Zampini 466552e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 466652e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 466752e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 466852e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4669b0c7d250SStefano Zampini if (!redprocs) { 4670b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 467128143c3dSStefano Zampini } else { 4672b0c7d250SStefano Zampini PetscInt idxs[1]; 4673b0c7d250SStefano Zampini PetscMPIInt tag; 4674b0c7d250SStefano Zampini MPI_Request *reqs; 4675b0c7d250SStefano Zampini 4676b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4677b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4678b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4679b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 468028143c3dSStefano Zampini } 4681b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4682b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4683b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4684b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4685e7931f94SStefano Zampini } 468652e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4687e7931f94SStefano Zampini /* clean up */ 4688e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 468952e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4690e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4691e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4692e7931f94SStefano Zampini } 4693e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4694e7931f94SStefano Zampini 4695e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4696e7931f94SStefano Zampini i = 1; 4697e7931f94SStefano Zampini if (color) i=0; 4698e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4699e7931f94SStefano Zampini /* get back IS */ 4700e7931f94SStefano Zampini *is_sends = ranks_send_to; 4701e7931f94SStefano Zampini PetscFunctionReturn(0); 4702e7931f94SStefano Zampini } 4703e7931f94SStefano Zampini 4704e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4705e7931f94SStefano Zampini 4706e7931f94SStefano Zampini #undef __FUNCT__ 4707e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 470853a05cb3SStefano 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[]) 4709e7931f94SStefano Zampini { 471070cf5478SStefano Zampini Mat local_mat; 4711e7931f94SStefano Zampini IS is_sends_internal; 47129d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 471328143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 47149d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4715e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4716e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4717e7931f94SStefano Zampini const PetscInt* is_indices; 4718e7931f94SStefano Zampini MatType new_local_type; 4719e7931f94SStefano Zampini /* buffers */ 4720e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 472128143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 47229d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4723e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4724e7931f94SStefano Zampini /* MPI */ 472528143c3dSStefano Zampini MPI_Comm comm,comm_n; 472628143c3dSStefano Zampini PetscSubcomm subcomm; 4727e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 472828143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 472928143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 473028143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 473128143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 473228143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4733e7931f94SStefano Zampini PetscErrorCode ierr; 4734e7931f94SStefano Zampini 4735e7931f94SStefano Zampini PetscFunctionBegin; 473628143c3dSStefano Zampini /* TODO: add missing checks */ 473728143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 473828143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 473928143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 474028143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4741e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 474228143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4743e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4744e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4745e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4746e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4747e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 474828143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 474970cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 475070cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 475128143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 475270cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 475370cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 475470cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 475570cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 475670cf5478SStefano Zampini } 4757e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4758e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4759e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4760e7931f94SStefano Zampini if (!is_sends) { 476128143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4762b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4763c8587f34SStefano Zampini } else { 4764e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4765e7931f94SStefano Zampini is_sends_internal = is_sends; 4766c8587f34SStefano Zampini } 4767e7931f94SStefano Zampini 4768e7931f94SStefano Zampini /* get comm */ 4769a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4770e7931f94SStefano Zampini 4771e7931f94SStefano Zampini /* compute number of sends */ 4772e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4773e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4774e7931f94SStefano Zampini 4775e7931f94SStefano Zampini /* compute number of receives */ 4776e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4777785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4778e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4779e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4780e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4781e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4782e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4783e7931f94SStefano Zampini 478428143c3dSStefano Zampini /* restrict comm if requested */ 478528143c3dSStefano Zampini subcomm = 0; 478628143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 478728143c3dSStefano Zampini if (restrict_comm) { 4788779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4789779c1cceSStefano Zampini 479028143c3dSStefano Zampini color = 0; 479153a05cb3SStefano Zampini if (restrict_full) { 479253a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 479353a05cb3SStefano Zampini } else { 479453a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 479553a05cb3SStefano Zampini } 479628143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 479728143c3dSStefano Zampini subcommsize = commsize - subcommsize; 479828143c3dSStefano Zampini /* check if reuse has been requested */ 479928143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 480028143c3dSStefano Zampini if (*mat_n) { 480128143c3dSStefano Zampini PetscMPIInt subcommsize2; 480228143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 480328143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 480428143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 480528143c3dSStefano Zampini } else { 480628143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 480728143c3dSStefano Zampini } 480828143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4809779c1cceSStefano Zampini PetscMPIInt rank; 4810779c1cceSStefano Zampini 4811779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 481228143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 481328143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 481428143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4815306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 481628143c3dSStefano Zampini } 481728143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 481828143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 481928143c3dSStefano Zampini } else { 482028143c3dSStefano Zampini comm_n = comm; 482128143c3dSStefano Zampini } 482228143c3dSStefano Zampini 4823e7931f94SStefano Zampini /* prepare send/receive buffers */ 4824785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4825e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4826785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4827e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 482828143c3dSStefano Zampini if (nis) { 4829854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 483028143c3dSStefano Zampini } 4831e7931f94SStefano Zampini 483228143c3dSStefano Zampini /* Get data from local matrices */ 4833e7931f94SStefano Zampini if (!isdense) { 4834a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4835e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4836e7931f94SStefano Zampini /* 4837e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4838e7931f94SStefano Zampini send_buffer_idxs should contain: 4839e7931f94SStefano Zampini - MatType_PRIVATE type 4840e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4841e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4842e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4843e7931f94SStefano Zampini */ 4844e7931f94SStefano Zampini } else { 4845e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 48463bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4847854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4848e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4849e7931f94SStefano Zampini send_buffer_idxs[1] = i; 48503bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4851e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 48523bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4853e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4854e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4855e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4856e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4857c8587f34SStefano Zampini } 4858c8587f34SStefano Zampini } 4859e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 486028143c3dSStefano Zampini /* additional is (if any) */ 486128143c3dSStefano Zampini if (nis) { 486228143c3dSStefano Zampini PetscMPIInt psum; 486328143c3dSStefano Zampini PetscInt j; 486428143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 486528143c3dSStefano Zampini PetscInt plen; 486628143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 486728143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 486828143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 486928143c3dSStefano Zampini } 4870854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 487128143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 487228143c3dSStefano Zampini PetscInt plen; 487328143c3dSStefano Zampini const PetscInt *is_array_idxs; 487428143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 487528143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 487628143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 487728143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 487828143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 487928143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 488028143c3dSStefano Zampini } 488128143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 488228143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 488328143c3dSStefano Zampini } 488428143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 488528143c3dSStefano Zampini } 488628143c3dSStefano Zampini 4887e7931f94SStefano Zampini buf_size_idxs = 0; 4888e7931f94SStefano Zampini buf_size_vals = 0; 488928143c3dSStefano Zampini buf_size_idxs_is = 0; 4890e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4891e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4892e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 489328143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4894e7931f94SStefano Zampini } 4895785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4896785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 489795ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4898e7931f94SStefano Zampini 4899e7931f94SStefano Zampini /* get new tags for clean communications */ 4900e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4901e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 490228143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4903e7931f94SStefano Zampini 4904e7931f94SStefano Zampini /* allocate for requests */ 4905785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4906785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 490795ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4908785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4909785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 491095ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4911e7931f94SStefano Zampini 4912e7931f94SStefano Zampini /* communications */ 4913e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4914e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 491528143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4916e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4917e7931f94SStefano Zampini source_dest = onodes[i]; 4918e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4919e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4920e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4921e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 492228143c3dSStefano Zampini if (nis) { 492328143c3dSStefano 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); 492428143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 492528143c3dSStefano Zampini } 4926e7931f94SStefano Zampini } 4927e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4928e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 4929e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 4930e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 493128143c3dSStefano Zampini if (nis) { 493228143c3dSStefano 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); 493328143c3dSStefano Zampini } 4934e7931f94SStefano Zampini } 4935e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4936e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 4937e7931f94SStefano Zampini 4938e7931f94SStefano Zampini /* assemble new l2g map */ 4939e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4940e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49419d30be91SStefano Zampini new_local_rows = 0; 4942e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49439d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4944e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4945e7931f94SStefano Zampini } 49469d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 4947e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 49489d30be91SStefano Zampini new_local_rows = 0; 4949e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 49509d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 49519d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 4952e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4953e7931f94SStefano Zampini } 49549d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 49559d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 4956e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 4957e7931f94SStefano Zampini 4958e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 4959e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 4960e7931f94SStefano 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) */ 4961e7931f94SStefano Zampini if (n_recvs) { 496228143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 4963e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4964e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4965e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 4966e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 4967e7931f94SStefano Zampini break; 4968e7931f94SStefano Zampini } 4969e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4970e7931f94SStefano Zampini } 4971e7931f94SStefano Zampini switch (new_local_type_private) { 497228143c3dSStefano Zampini case MATDENSE_PRIVATE: 497328143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 4974e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4975e7931f94SStefano Zampini bs = 1; 497628143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 497728143c3dSStefano Zampini new_local_type = MATSEQDENSE; 497828143c3dSStefano Zampini bs = 1; 497928143c3dSStefano Zampini } 4980e7931f94SStefano Zampini break; 4981e7931f94SStefano Zampini case MATAIJ_PRIVATE: 4982e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 4983e7931f94SStefano Zampini bs = 1; 4984e7931f94SStefano Zampini break; 4985e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 4986e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 4987e7931f94SStefano Zampini break; 4988e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 4989e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 4990e7931f94SStefano Zampini break; 4991e7931f94SStefano Zampini default: 49929d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 4993e7931f94SStefano Zampini break; 4994e7931f94SStefano Zampini } 499528143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 499628143c3dSStefano Zampini new_local_type = MATSEQDENSE; 499728143c3dSStefano Zampini bs = 1; 4998e7931f94SStefano Zampini } 4999e7931f94SStefano Zampini 500070cf5478SStefano Zampini /* create MATIS object if needed */ 500170cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5002e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5003e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 500470cf5478SStefano Zampini } else { 500570cf5478SStefano Zampini /* it also destroys the local matrices */ 500670cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 500770cf5478SStefano Zampini } 500870cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5009e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 50109d30be91SStefano Zampini 50119d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 50129d30be91SStefano Zampini 50139d30be91SStefano Zampini /* Global to local map of received indices */ 50149d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 50159d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 50169d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 50179d30be91SStefano Zampini 50189d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 50199d30be91SStefano Zampini buf_size_idxs = 0; 50209d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50219d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 50229d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 50239d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 50249d30be91SStefano Zampini } 50259d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 50269d30be91SStefano Zampini 50279d30be91SStefano Zampini /* set preallocation */ 50289d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 50299d30be91SStefano Zampini if (!newisdense) { 50309d30be91SStefano Zampini PetscInt *new_local_nnz=0; 50319d30be91SStefano Zampini 50329d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 50339d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 50349d30be91SStefano Zampini if (n_recvs) { 50359d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 50369d30be91SStefano Zampini } 50379d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50389d30be91SStefano Zampini PetscInt j; 50399d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 50409d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 50419d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 50429d30be91SStefano Zampini } 50439d30be91SStefano Zampini } else { 50449d30be91SStefano Zampini /* TODO */ 50459d30be91SStefano Zampini } 50469d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 50479d30be91SStefano Zampini } 50489d30be91SStefano Zampini if (new_local_nnz) { 50499d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 50509d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 50519d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 50529d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50539d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 50549d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 50559d30be91SStefano Zampini } else { 50569d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50579d30be91SStefano Zampini } 50589d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 50599d30be91SStefano Zampini } else { 50609d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 50619d30be91SStefano Zampini } 5062e7931f94SStefano Zampini 5063e7931f94SStefano Zampini /* set values */ 5064e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 50659d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5066e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5067e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5068e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 50699d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5070e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5071e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5072e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 507328143c3dSStefano Zampini } else { 507428143c3dSStefano Zampini /* TODO */ 5075e7931f94SStefano Zampini } 5076e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5077e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5078e7931f94SStefano Zampini } 5079e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5080e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 508170cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 508270cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 50839d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 50849d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5085e7931f94SStefano Zampini 5086dfd14d43SStefano Zampini #if 0 508728143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5088e7931f94SStefano Zampini Vec lvec,rvec; 5089e7931f94SStefano Zampini PetscReal infty_error; 5090e7931f94SStefano Zampini 50912a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5092e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5093e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5094e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 509570cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5096e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5097e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5098e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5099e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5100e7931f94SStefano Zampini } 510128143c3dSStefano Zampini #endif 5102e7931f94SStefano Zampini 510328143c3dSStefano Zampini /* assemble new additional is (if any) */ 510428143c3dSStefano Zampini if (nis) { 510528143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 510628143c3dSStefano Zampini 510728143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5108854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 510928143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 511028143c3dSStefano Zampini psum = 0; 511128143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 511228143c3dSStefano Zampini for (j=0;j<nis;j++) { 511328143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 511428143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 511528143c3dSStefano Zampini psum += plen; 511628143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 511728143c3dSStefano Zampini } 511828143c3dSStefano Zampini } 5119854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5120854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 512128143c3dSStefano Zampini for (i=1;i<nis;i++) { 512228143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 512328143c3dSStefano Zampini } 512428143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 512528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 512628143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 512728143c3dSStefano Zampini for (j=0;j<nis;j++) { 512828143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 512928143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 513028143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 513128143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 513228143c3dSStefano Zampini } 513328143c3dSStefano Zampini } 513428143c3dSStefano Zampini for (i=0;i<nis;i++) { 513528143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 513628143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 513728143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 513828143c3dSStefano Zampini } 513928143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 514028143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 514128143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 514228143c3dSStefano Zampini } 5143e7931f94SStefano Zampini /* free workspace */ 514428143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5145e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5146e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5147e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5148e7931f94SStefano Zampini if (isdense) { 5149e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5150e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5151e7931f94SStefano Zampini } else { 5152e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5153e7931f94SStefano Zampini } 515428143c3dSStefano Zampini if (nis) { 515528143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 515628143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 515728143c3dSStefano Zampini } 5158e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5159e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 516028143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5161e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5162e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 516328143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5164e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5165e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5166e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5167e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5168e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 516928143c3dSStefano Zampini if (nis) { 517028143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 517128143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 517228143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 517328143c3dSStefano Zampini } 517428143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 517528143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 517628143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 517728143c3dSStefano Zampini for (i=0;i<nis;i++) { 517828143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 517928143c3dSStefano Zampini } 518053a05cb3SStefano Zampini *mat_n = NULL; 518128143c3dSStefano Zampini } 5182e7931f94SStefano Zampini PetscFunctionReturn(0); 5183e7931f94SStefano Zampini } 5184a57a6d2fSStefano Zampini 518512edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5186af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 518712edc857SStefano Zampini 5188c8587f34SStefano Zampini #undef __FUNCT__ 5189c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5190c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5191c8587f34SStefano Zampini { 5192c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5193c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 519420a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 51959881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 519620a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 51976e683305SStefano Zampini IS coarse_is,*isarray; 51986e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 519930368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5200f9eb5b7dSStefano Zampini PC pc_temp; 5201c8587f34SStefano Zampini PCType coarse_pc_type; 5202c8587f34SStefano Zampini KSPType coarse_ksp_type; 5203f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 52044f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 52056e683305SStefano Zampini Mat t_coarse_mat_is; 52066e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 52076e683305SStefano Zampini PetscMPIInt all_procs; 520874e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 520968457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 521022bc73bbSStefano Zampini PetscScalar *array; 52119881197aSStefano Zampini PetscErrorCode ierr; 5212fdc09c96SStefano Zampini 5213c8587f34SStefano Zampini PetscFunctionBegin; 5214c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 521568457ee5SStefano 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 */ 5216fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 52175a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5218fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5219f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5220f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5221f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5222fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 522351bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 522451bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5225dc4bcba2SStefano Zampini PC pc; 5226dc4bcba2SStefano Zampini PetscBool isbddc; 5227dc4bcba2SStefano Zampini 5228dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5229dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5230dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5231dc4bcba2SStefano Zampini if (isbddc) { 523263c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 523363c961adSStefano Zampini } else { 5234727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 523563c961adSStefano Zampini } 5236fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5237fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5238fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5239f4ddd8eeSStefano Zampini } 5240fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5241fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5242f4ddd8eeSStefano Zampini } 524370cf5478SStefano Zampini /* reset any subassembling information */ 524470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 52456e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 52466e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5247fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5248f4ddd8eeSStefano Zampini } 5249c8587f34SStefano Zampini 52506e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 52512b510759SStefano Zampini im_active = !!(pcis->n); 52522b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 52536e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 52546e683305SStefano Zampini void_procs = all_procs-active_procs; 52556e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 525674e2c79eSStefano Zampini redist = PETSC_FALSE; 525722bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 52586e683305SStefano Zampini csin_ml = PETSC_TRUE; 52596e683305SStefano Zampini ncoarse_ml = void_procs; 5260779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5261779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 52626e683305SStefano Zampini csin_ds = PETSC_TRUE; 526318a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 526418a45a71SStefano Zampini redist = PETSC_TRUE; 526518a45a71SStefano Zampini } else { 52666e683305SStefano Zampini csin_ds = PETSC_TRUE; 5267779c1cceSStefano Zampini ncoarse_ds = active_procs; 5268779c1cceSStefano Zampini redist = PETSC_TRUE; 526918a45a71SStefano Zampini } 52706e683305SStefano Zampini } else { 52716e683305SStefano Zampini csin_ml = PETSC_FALSE; 52726e683305SStefano Zampini ncoarse_ml = all_procs; 52736e683305SStefano Zampini if (void_procs) { 52746e683305SStefano Zampini csin_ds = PETSC_TRUE; 52756e683305SStefano Zampini ncoarse_ds = void_procs; 52766e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 52776e683305SStefano Zampini } else { 5278779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 527974e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 528074e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 528174e2c79eSStefano Zampini redist = PETSC_TRUE; 528274e2c79eSStefano Zampini } else { 52836e683305SStefano Zampini csin_ds = PETSC_FALSE; 52846e683305SStefano Zampini ncoarse_ds = all_procs; 52856e683305SStefano Zampini } 52866e683305SStefano Zampini } 528774e2c79eSStefano Zampini } 52886e683305SStefano Zampini 52896e683305SStefano Zampini /* 52906e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 52916e683305SStefano Zampini - we have not exceeded the number of levels requested 52926e683305SStefano Zampini - we can actually subassemble the active processes 52936e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 52946e683305SStefano Zampini */ 52956e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 52966e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 52976e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 52986e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 52996e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5300f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 53012b510759SStefano Zampini } else { 5302f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5303c8587f34SStefano Zampini } 5304c8587f34SStefano Zampini } 53056e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 53066e683305SStefano Zampini if (multilevel_allowed) { 53076e683305SStefano Zampini ncoarse = ncoarse_ml; 53086e683305SStefano Zampini csin = csin_ml; 530958da7f69SStefano Zampini redist = PETSC_FALSE; 53106e683305SStefano Zampini } else { 53116e683305SStefano Zampini ncoarse = ncoarse_ds; 53126e683305SStefano Zampini csin = csin_ds; 53136e683305SStefano Zampini } 5314e7931f94SStefano Zampini 5315abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5316abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5317abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5318abbbba34SStefano Zampini 5319abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 532022bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 532122bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 532222bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 532322bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5324e176bc59SStefano 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); 53256e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 53266e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 53276e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5328abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5329abbbba34SStefano Zampini 53306e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 533130368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 53326e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 53336e683305SStefano Zampini const PetscInt *idxs; 53346e683305SStefano Zampini ISLocalToGlobalMapping tmap; 53356e683305SStefano Zampini 53366e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 53370be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 53386e683305SStefano Zampini /* allocate space for temporary storage */ 5339854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5340854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 53416e683305SStefano Zampini /* allocate for IS array */ 53426e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 53436e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 534430368db7SStefano Zampini nisvert = 0; 534530368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 534630368db7SStefano Zampini nisvert = 1; 534730368db7SStefano Zampini } 534830368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5349854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 53506e683305SStefano Zampini /* dofs splitting */ 53516e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 53526e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 53536e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 53546e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53556e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53566e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 53576e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 535830368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 53596e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 53606e683305SStefano Zampini } 53616e683305SStefano Zampini /* neumann boundaries */ 53626e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 53636e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 53646e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 53656e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53666e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 53676e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 53686e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 536930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 53706e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 53716e683305SStefano Zampini } 537230368db7SStefano Zampini /* primal vertices (benign) */ 537330368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 537430368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 537530368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 537630368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 537730368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 537830368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 537930368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 538030368db7SStefano Zampini } 53816e683305SStefano Zampini /* free memory */ 53826e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 53836e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 53846e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 53856e683305SStefano Zampini } else { 53866e683305SStefano Zampini nis = 0; 53876e683305SStefano Zampini nisdofs = 0; 53886e683305SStefano Zampini nisneu = 0; 538930368db7SStefano Zampini nisvert = 0; 53906e683305SStefano Zampini isarray = NULL; 53916e683305SStefano Zampini } 53926e683305SStefano Zampini /* destroy no longer needed map */ 53936e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 53946e683305SStefano Zampini 53956e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 53966e683305SStefano Zampini coarse_mat_is = NULL; 53976e683305SStefano Zampini if (csin) { 53986e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 539974e2c79eSStefano Zampini if (redist) { 540074e2c79eSStefano Zampini PetscMPIInt rank; 5401779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 540274e2c79eSStefano Zampini 540374e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 540458da7f69SStefano Zampini spc = active_procs/ncoarse; 540558da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5406779c1cceSStefano Zampini if (im_active) { 5407779c1cceSStefano Zampini destsize = 1; 540874e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 540974e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 541074e2c79eSStefano Zampini } else { 541174e2c79eSStefano Zampini dest[0] = rank/(spc+1); 541274e2c79eSStefano Zampini } 541374e2c79eSStefano Zampini } else { 5414779c1cceSStefano Zampini destsize = 0; 54156e683305SStefano Zampini } 5416779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5417779c1cceSStefano Zampini } else if (csin_type_simple) { 54186e683305SStefano Zampini PetscMPIInt rank; 54196e683305SStefano Zampini PetscInt issize,isidx; 5420779c1cceSStefano Zampini 54216e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 54226e683305SStefano Zampini if (im_active) { 54236e683305SStefano Zampini issize = 1; 54246e683305SStefano Zampini isidx = (PetscInt)rank; 54256e683305SStefano Zampini } else { 54266e683305SStefano Zampini issize = 0; 54276e683305SStefano Zampini isidx = -1; 54286e683305SStefano Zampini } 54296e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5430779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5431b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 54326e683305SStefano Zampini } 5433779c1cceSStefano Zampini 5434779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5435779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5436779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5437779c1cceSStefano Zampini PetscInt *coarse_candidates; 5438779c1cceSStefano Zampini const PetscInt* tisindices; 5439779c1cceSStefano Zampini 5440779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5441779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5442779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5443779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5444779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5445779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5446779c1cceSStefano Zampini } 5447779c1cceSStefano Zampini } 5448779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5449779c1cceSStefano Zampini 5450779c1cceSStefano Zampini 54516e683305SStefano Zampini if (pcbddc->dbg_flag) { 54526e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54536e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 54546e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54556e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5456779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 54576e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 54586e683305SStefano Zampini } 54596e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 54606e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54616e683305SStefano Zampini } 54626e683305SStefano Zampini /* shift the pattern on coarse candidates */ 54636e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 54646e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5465854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 54666e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 54676e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 54686e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 54696e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 54706e683305SStefano Zampini } 54716e683305SStefano Zampini if (pcbddc->dbg_flag) { 54726e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54736e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 54746e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 54756e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54766e683305SStefano Zampini } 5477779c1cceSStefano Zampini } 54786e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 547953a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 548053a05cb3SStefano 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); 548153a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 548253a05cb3SStefano 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); 548353a05cb3SStefano Zampini } 54846e683305SStefano Zampini } else { 54856e683305SStefano Zampini if (pcbddc->dbg_flag) { 54866e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 54876e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 54886e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 54896e683305SStefano Zampini } 54906e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 54916e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 54926e683305SStefano Zampini } 54936e683305SStefano Zampini 54946e683305SStefano Zampini /* create local to global scatters for coarse problem */ 549568457ee5SStefano Zampini if (compute_vecs) { 54966e683305SStefano Zampini PetscInt lrows; 54976e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 54986e683305SStefano Zampini if (coarse_mat_is) { 54996e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 55006e683305SStefano Zampini } else { 55016e683305SStefano Zampini lrows = 0; 55026e683305SStefano Zampini } 55036e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 55046e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 55056e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 55066e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55076e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55086e683305SStefano Zampini } 55096e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 55106e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5511c8587f34SStefano Zampini 5512f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5513f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5514f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5515f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5516f9eb5b7dSStefano Zampini } else { 5517f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5518f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5519c8587f34SStefano Zampini } 5520c8587f34SStefano Zampini 55216e683305SStefano Zampini /* print some info if requested */ 55226e683305SStefano Zampini if (pcbddc->dbg_flag) { 55236e683305SStefano Zampini if (!multilevel_allowed) { 55246e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55256e683305SStefano Zampini if (multilevel_requested) { 55266e683305SStefano 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); 55276e683305SStefano Zampini } else if (pcbddc->max_levels) { 55286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 55296e683305SStefano Zampini } 55306e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55316e683305SStefano Zampini } 55326e683305SStefano Zampini } 55336e683305SStefano Zampini 5534f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 55356e683305SStefano Zampini if (coarse_mat_is) { 55366e683305SStefano Zampini MatReuse coarse_mat_reuse; 55376a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 55386e683305SStefano Zampini if (pcbddc->dbg_flag) { 55396e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 55406e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 55416e683305SStefano Zampini } 5542f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5543312be037SStefano Zampini char prefix[256],str_level[16]; 5544e604994aSStefano Zampini size_t len; 55456e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5546422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5547c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5548f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 55495f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5550c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 55516e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5552c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5553c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5554e604994aSStefano Zampini /* prefix */ 5555e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5556e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5557e604994aSStefano Zampini if (!pcbddc->current_level) { 5558e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5559e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5560c8587f34SStefano Zampini } else { 5561e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5562312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5563312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 556434d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5565312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5566e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5567e604994aSStefano Zampini } 5568e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 55693e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 55703e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 55713e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 55723e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5573f9eb5b7dSStefano Zampini /* allow user customization */ 5574f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 55753e3c6dadSStefano Zampini } 55763e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 557751bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 55783e3c6dadSStefano Zampini if (nisdofs) { 55793e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 55803e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 55813e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 55823e3c6dadSStefano Zampini } 55833e3c6dadSStefano Zampini } 55843e3c6dadSStefano Zampini if (nisneu) { 55853e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 55863e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5587312be037SStefano Zampini } 558830368db7SStefano Zampini if (nisvert) { 558930368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 559030368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 559130368db7SStefano Zampini } 5592f9eb5b7dSStefano Zampini 5593f9eb5b7dSStefano Zampini /* get some info after set from options */ 5594f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5595f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 55964f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 55976e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5598f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5599f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5600f9eb5b7dSStefano Zampini } 560139f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 56024f3a063dSStefano Zampini if (isredundant) { 56034f3a063dSStefano Zampini KSP inner_ksp; 56044f3a063dSStefano Zampini PC inner_pc; 56054f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 56064f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 56074f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 56084f3a063dSStefano Zampini } 5609f9eb5b7dSStefano Zampini 5610f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5611fa7f1dd8SStefano Zampini if (coarse_reuse) { 561281d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5613fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 56146e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5615fa7f1dd8SStefano Zampini } else { 56166e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5617fa7f1dd8SStefano Zampini } 5618c8587f34SStefano Zampini if (isbddc || isnn) { 5619d4d8cf7bSStefano Zampini if (isbddc) { /* currently there's no API for this */ 5620d4d8cf7bSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc_temp->data; 5621d4d8cf7bSStefano Zampini pcbddc->detect_disconnected = PETSC_TRUE; 5622d4d8cf7bSStefano Zampini } 562322bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 562470cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5625b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 562622b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 56276e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 56296e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 56306e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 563122b6e8a2SStefano Zampini } 563270cf5478SStefano Zampini } 563353a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 563470cf5478SStefano Zampini } else { 563522bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 563622bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 563722bc73bbSStefano Zampini } 563822bc73bbSStefano Zampini } else { 56392e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5640c8587f34SStefano Zampini } 5641c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5642c8587f34SStefano Zampini 56433301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 56445a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 56453301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 56463301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 56473301b35fSStefano Zampini } 56483301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 56493301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 56503301b35fSStefano Zampini } 56513301b35fSStefano Zampini if (pc->pmat->spd_set) { 56523301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 56533301b35fSStefano Zampini } 56546e683305SStefano Zampini /* set operators */ 56555f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 56566e683305SStefano Zampini if (pcbddc->dbg_flag) { 56576e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56586e683305SStefano Zampini } 56596e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 56606e683305SStefano Zampini coarse_mat = 0; 56616e683305SStefano Zampini } 56626e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5663b1ecc7b1SStefano Zampini #if 0 5664b9b85e73SStefano Zampini { 5665b9b85e73SStefano Zampini PetscViewer viewer; 5666b9b85e73SStefano Zampini char filename[256]; 5667b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5668b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5669b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5670b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5671b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5672b9b85e73SStefano Zampini } 5673b9b85e73SStefano Zampini #endif 5674c8587f34SStefano Zampini 5675c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5676298c0119SStefano Zampini #if 0 5677c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5678c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 567998a51de6SStefano Zampini } 5680298c0119SStefano Zampini #endif 5681b0f5fe93SStefano Zampini /* hack */ 568298a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 568398a51de6SStefano Zampini Vec crhs,csol; 568404708bb6SStefano Zampini 5685f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5686f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5687f347579bSStefano Zampini if (!csol) { 56882a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5689f9eb5b7dSStefano Zampini } 5690f347579bSStefano Zampini if (!crhs) { 56912a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5692f347579bSStefano Zampini } 5693b0f5fe93SStefano Zampini } 5694b0f5fe93SStefano Zampini 5695b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5696b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5697b0f5fe93SStefano Zampini 5698b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 56994f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 57004f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 57014f1b2e48SStefano Zampini } 5702b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5703b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5704b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5705b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5706b0f5fe93SStefano Zampini if (coarse_mat) { 5707b0f5fe93SStefano Zampini Vec nullv; 5708b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5709b0f5fe93SStefano Zampini PetscInt nl; 5710b0f5fe93SStefano Zampini 5711b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5712b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5713b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5714b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5715b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5716b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5717b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5718b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5719b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5720b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5721b0f5fe93SStefano Zampini } 5722b0f5fe93SStefano Zampini } 5723b0f5fe93SStefano Zampini 5724b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5725b0f5fe93SStefano Zampini PetscBool ispreonly; 5726b0f5fe93SStefano Zampini 5727b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5728b0f5fe93SStefano Zampini PetscBool isnull; 5729b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5730d4d8cf7bSStefano Zampini if (0) { 573130368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5732b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5733b0f5fe93SStefano Zampini } else { 5734b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5735b0f5fe93SStefano Zampini } 5736b0f5fe93SStefano Zampini } else { 5737b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5738b0f5fe93SStefano Zampini } 5739b0f5fe93SStefano Zampini } 5740b0f5fe93SStefano Zampini /* setup coarse ksp */ 5741b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5742cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5743cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 57446e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5745c8587f34SStefano Zampini KSP check_ksp; 57462b510759SStefano Zampini KSPType check_ksp_type; 5747c8587f34SStefano Zampini PC check_pc; 57486e683305SStefano Zampini Vec check_vec,coarse_vec; 57496a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 57502b510759SStefano Zampini PetscInt its; 57516e683305SStefano Zampini PetscBool compute_eigs; 57526e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 57536e683305SStefano Zampini PetscInt neigs; 57548e185a42SStefano Zampini const char *prefix; 5755c8587f34SStefano Zampini 57562b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 57576e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5758422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 575923ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5760f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 57612b510759SStefano Zampini if (ispreonly) { 57622b510759SStefano Zampini check_ksp_type = KSPPREONLY; 57636e683305SStefano Zampini compute_eigs = PETSC_FALSE; 57642b510759SStefano Zampini } else { 5765cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 57666e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5767c8587f34SStefano Zampini } 5768c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 57696e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 57706e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 57716e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5772a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5773a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5774a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5775a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5776c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5777c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5778c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5779c8587f34SStefano Zampini /* create random vec */ 57806e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 57816e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5782c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5783c8587f34SStefano Zampini if (CoarseNullSpace) { 5784c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5785c8587f34SStefano Zampini } 57866e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5787c8587f34SStefano Zampini /* solve coarse problem */ 57886e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5789c8587f34SStefano Zampini if (CoarseNullSpace) { 57906e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5791c8587f34SStefano Zampini } 5792cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 57936e683305SStefano Zampini if (compute_eigs) { 5794854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5795854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 57966e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 57976e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 57986e683305SStefano Zampini lambda_min = eigs_r[0]; 57996e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 58006e683305SStefano Zampini if (lambda_max>lambda_min) { 5801cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5802cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5803cbcc2c2aSStefano Zampini } 5804c8587f34SStefano Zampini } 5805c8587f34SStefano Zampini } 5806cbcc2c2aSStefano Zampini 5807c8587f34SStefano Zampini /* check coarse problem residual error */ 58086e683305SStefano Zampini if (pcbddc->dbg_flag) { 58096e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 58106e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58116e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5812c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 58136e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 58146e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5815c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5816779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 58176e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 58186e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 58196e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 58206e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5821b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5822b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5823b0f5fe93SStefano Zampini } 58246e683305SStefano Zampini if (compute_eigs) { 58256e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5826deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5827c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 58286e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 58296e683305SStefano 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); 58306e683305SStefano Zampini for (i=0;i<neigs;i++) { 58316e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5832c8587f34SStefano Zampini } 58336e683305SStefano Zampini } 58346e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 58356e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58366e683305SStefano Zampini } 5837c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 58386e683305SStefano Zampini if (compute_eigs) { 58396e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 58406e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5841c8587f34SStefano Zampini } 58426e683305SStefano Zampini } 58436e683305SStefano Zampini } 5844cbcc2c2aSStefano Zampini /* print additional info */ 5845cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 58466e683305SStefano Zampini /* waits until all processes reaches this point */ 58476e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5848cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5849cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5850cbcc2c2aSStefano Zampini } 5851cbcc2c2aSStefano Zampini 58522b510759SStefano Zampini /* free memory */ 5853c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5854fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5855c8587f34SStefano Zampini PetscFunctionReturn(0); 5856c8587f34SStefano Zampini } 5857674ae819SStefano Zampini 5858f34684f1SStefano Zampini #undef __FUNCT__ 5859f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5860f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5861f34684f1SStefano Zampini { 5862f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5863f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5864f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5865dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5866dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 586773be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5868dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5869f34684f1SStefano Zampini PetscErrorCode ierr; 5870f34684f1SStefano Zampini 5871f34684f1SStefano Zampini PetscFunctionBegin; 5872f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 58730e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 58740e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5875727cdba6SStefano Zampini } 5876dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 58773bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5878dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5879dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5880dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5881dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5882dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5883dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 58840e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 58850e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 58860e6343abSStefano Zampini } 5887dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5888dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5889dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5890dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5891dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5892f34684f1SStefano Zampini 5893f34684f1SStefano Zampini /* check numbering */ 5894f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5895019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5896dc456d91SStefano Zampini PetscInt i; 5897b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5898f34684f1SStefano Zampini 5899f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5900f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5901f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 59021575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5903019a44ceSStefano Zampini /* counter */ 5904019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5905019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5906019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5907019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5908019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5909019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5910f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5911f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5912727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5913f34684f1SStefano Zampini } 5914f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5915f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5916f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5917e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5918e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5919e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5920e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5921f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5922019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5923f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5924019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 592575c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 592675c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 5927b9b85e73SStefano Zampini set_error = PETSC_TRUE; 5928019a44ceSStefano 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); 5929f34684f1SStefano Zampini } 5930f34684f1SStefano Zampini } 5931019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5932b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5933f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5934f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 5935f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 5936f34684f1SStefano Zampini } 5937f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5938f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5939e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5940e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5941f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 5942f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 5943b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 5944ca8b9ea9SStefano Zampini PetscInt *gidxs; 5945ca8b9ea9SStefano Zampini 5946ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 59473bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 5948f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 5949f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5950f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 5951f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 59524bc2dc4bSStefano 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); 5953f34684f1SStefano Zampini } 5954f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5955ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 5956f34684f1SStefano Zampini } 5957f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 59581575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5959302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 5960f34684f1SStefano Zampini } 59618bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 5962f34684f1SStefano Zampini /* get back data */ 5963f34684f1SStefano Zampini *coarse_size_n = coarse_size; 5964f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 5965674ae819SStefano Zampini PetscFunctionReturn(0); 5966674ae819SStefano Zampini } 5967674ae819SStefano Zampini 5968e456f2a8SStefano Zampini #undef __FUNCT__ 5969e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 5970a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 5971e456f2a8SStefano Zampini { 5972e456f2a8SStefano Zampini IS localis_t; 5973a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 5974e456f2a8SStefano Zampini PetscScalar *vals; 5975e456f2a8SStefano Zampini PetscErrorCode ierr; 5976e456f2a8SStefano Zampini 5977e456f2a8SStefano Zampini PetscFunctionBegin; 5978a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 5979e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 5980854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 5981e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 5982e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5983a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 5984a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 59851035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 5986a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 59871035eff8SStefano Zampini } 5988a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 5989e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 5990e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 5991a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 5992a7dc3881SStefano Zampini /* now compute set in local ordering */ 5993a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5994a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5995a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 5996a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 5997a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 5998ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 5999e456f2a8SStefano Zampini lsize++; 6000e456f2a8SStefano Zampini } 6001e456f2a8SStefano Zampini } 6002854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6003a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6004ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6005e456f2a8SStefano Zampini idxs[lsize++] = i; 6006e456f2a8SStefano Zampini } 6007e456f2a8SStefano Zampini } 6008a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6009a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6010e456f2a8SStefano Zampini *localis = localis_t; 6011e456f2a8SStefano Zampini PetscFunctionReturn(0); 6012e456f2a8SStefano Zampini } 6013906d46d4SStefano Zampini 6014906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6015906d46d4SStefano Zampini #undef __FUNCT__ 6016906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6017906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6018906d46d4SStefano Zampini { 6019906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6020906d46d4SStefano Zampini PetscErrorCode ierr; 6021906d46d4SStefano Zampini 6022906d46d4SStefano Zampini PetscFunctionBegin; 6023906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6024906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6025906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6026906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6027906d46d4SStefano Zampini PetscFunctionReturn(0); 6028906d46d4SStefano Zampini } 6029906d46d4SStefano Zampini 6030906d46d4SStefano Zampini #undef __FUNCT__ 6031906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6032906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6033906d46d4SStefano Zampini { 6034906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6035906d46d4SStefano Zampini PetscErrorCode ierr; 6036906d46d4SStefano Zampini 6037906d46d4SStefano Zampini PetscFunctionBegin; 6038906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6039906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6040906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6041906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6042906d46d4SStefano Zampini PetscFunctionReturn(0); 6043906d46d4SStefano Zampini } 6044b96c3477SStefano Zampini 6045b96c3477SStefano Zampini #undef __FUNCT__ 6046b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 604708122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6048b96c3477SStefano Zampini { 6049a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6050b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6051b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6052a64f4aa4SStefano Zampini Mat S_j; 6053b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6054b96c3477SStefano Zampini PetscBool free_used_adj; 6055b96c3477SStefano Zampini PetscErrorCode ierr; 6056b96c3477SStefano Zampini 6057b96c3477SStefano Zampini PetscFunctionBegin; 6058b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6059b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 606008122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6061b96c3477SStefano Zampini used_xadj = NULL; 6062b96c3477SStefano Zampini used_adjncy = NULL; 6063b96c3477SStefano Zampini } else { 606408122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 606508122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 606608122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 606708122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6068b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6069b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6070b96c3477SStefano Zampini } else { 60712fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6072b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6073b96c3477SStefano Zampini PetscInt nvtxs; 6074b96c3477SStefano Zampini 60752fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 60762fffb893SStefano Zampini if (flg_row) { 6077b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6078b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6079b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6080b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 60812fffb893SStefano Zampini } else { 60822fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 60832fffb893SStefano Zampini used_xadj = NULL; 60842fffb893SStefano Zampini used_adjncy = NULL; 60852fffb893SStefano Zampini } 60862fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6087b96c3477SStefano Zampini } 6088b96c3477SStefano Zampini } 6089d5574798SStefano Zampini 6090d5574798SStefano Zampini /* setup sub_schurs data */ 6091a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6092df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6093df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6094a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 609572b8c272SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 6096a64f4aa4SStefano Zampini } else { 60976816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 609804708bb6SStefano Zampini PetscBool isseqaij; 6099a3df083aSStefano Zampini PetscInt benign_n; 610072b8c272SStefano Zampini Mat change = NULL; 610172b8c272SStefano Zampini IS change_primal = NULL; 6102a3df083aSStefano Zampini 61035feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 61045feab87aSStefano Zampini PetscInt n_vertices; 61055feab87aSStefano Zampini 61065feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 61072034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 61085feab87aSStefano Zampini } 610904708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 611004708bb6SStefano Zampini if (!isseqaij) { 611104708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 611204708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 611304708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 611404708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 611504708bb6SStefano Zampini } else { 611604708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 611704708bb6SStefano Zampini } 611804708bb6SStefano Zampini } 6119a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6120a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6121ca92afb2SStefano Zampini } else { 6122a3df083aSStefano Zampini benign_n = 0; 6123ca92afb2SStefano Zampini } 6124bf3a8328SStefano Zampini /* If the user defines additional constraints, we import them here. 6125bf3a8328SStefano Zampini We need to compute the change of basis according to the quadrature weights attached to pmat via MatSetNearNullSpace, and this could not be done, at the moment, without some hacking */ 612672b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 612788c03ad3SStefano Zampini PC_IS *pcisf; 612888c03ad3SStefano Zampini PC_BDDC *pcbddcf; 612988c03ad3SStefano Zampini PC pcf; 613088c03ad3SStefano Zampini 613188c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 613288c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 613388c03ad3SStefano Zampini } 613488c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 613588c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 613688c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 613788c03ad3SStefano Zampini /* hacks */ 613888c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 613972b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 614072b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 614172b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 614272b8c272SStefano Zampini pcisf->n = pcis->n; 614372b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 614488c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 614588c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 614688c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 614788c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 614888c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 614988c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 615072b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 615188c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 615288c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 615372b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 615472b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 615572b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 615672b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 615772b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 615888c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 615972b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 616088c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 616188c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 616288c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 616388c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 616488c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 616588c03ad3SStefano Zampini pcf->ops->destroy = NULL; 616688c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 616788c03ad3SStefano Zampini } 616872b8c272SStefano Zampini ierr = PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,pcbddc->faster_deluxe,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);CHKERRQ(ierr); 616972b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 617072b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6171ca92afb2SStefano Zampini } 6172d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6173b96c3477SStefano Zampini 6174b96c3477SStefano Zampini /* free adjacency */ 6175b96c3477SStefano Zampini if (free_used_adj) { 6176b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6177b96c3477SStefano Zampini } 6178b96c3477SStefano Zampini PetscFunctionReturn(0); 6179b96c3477SStefano Zampini } 6180b96c3477SStefano Zampini 6181b96c3477SStefano Zampini #undef __FUNCT__ 6182b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 618308122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6184b96c3477SStefano Zampini { 6185b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6186b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6187b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6188b96c3477SStefano Zampini PCBDDCGraph graph; 6189b96c3477SStefano Zampini PetscErrorCode ierr; 6190b96c3477SStefano Zampini 6191b96c3477SStefano Zampini PetscFunctionBegin; 6192b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 619308122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 61943301b35fSStefano Zampini IS verticesIS,verticescomm; 61953301b35fSStefano Zampini PetscInt vsize,*idxs; 6196b96c3477SStefano Zampini 6197b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 61983301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 61993301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62003301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 62013301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62023301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6203b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 62047fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 62053301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 62063301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6207b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6208b96c3477SStefano Zampini /* 6209b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6210b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6211b96c3477SStefano Zampini } 6212b96c3477SStefano Zampini */ 6213b96c3477SStefano Zampini } else { 6214b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6215b96c3477SStefano Zampini } 6216b96c3477SStefano Zampini 6217b96c3477SStefano Zampini /* sub_schurs init */ 6218a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6219a64f4aa4SStefano Zampini 6220b96c3477SStefano Zampini /* free graph struct */ 622108122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6222b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6223b96c3477SStefano Zampini } 6224b96c3477SStefano Zampini PetscFunctionReturn(0); 6225b96c3477SStefano Zampini } 6226fa34dd3eSStefano Zampini 6227fa34dd3eSStefano Zampini #undef __FUNCT__ 6228fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6229fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6230fa34dd3eSStefano Zampini { 6231fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6232fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6233fa34dd3eSStefano Zampini PetscErrorCode ierr; 6234fa34dd3eSStefano Zampini 6235fa34dd3eSStefano Zampini PetscFunctionBegin; 6236fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6237fa34dd3eSStefano Zampini IS zerodiag = NULL; 62384f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6239fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 62404f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 624175c01103SStefano Zampini PetscReal norm; 6242fa34dd3eSStefano Zampini PetscInt i; 6243fa34dd3eSStefano Zampini 6244fa34dd3eSStefano Zampini /* B0 and B0_B */ 6245fa34dd3eSStefano Zampini if (zerodiag) { 6246fa34dd3eSStefano Zampini IS dummy; 6247fa34dd3eSStefano Zampini 62484f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 62494f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6250fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6251fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6252fa34dd3eSStefano Zampini } 6253fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6254fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6255fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6256fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6257fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6258fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6259fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6260fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6261fa34dd3eSStefano Zampini /* S_j */ 6262fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6263fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6264fa34dd3eSStefano Zampini 6265fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6266fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6267fa34dd3eSStefano Zampini /* continuous in primal space */ 6268fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6269fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6270fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6271fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 62724f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 62734f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6274fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6275fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6276fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6277fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6278fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6279fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6280fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6281fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6282fa34dd3eSStefano Zampini 6283fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6284fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6285fa34dd3eSStefano Zampini /* local with Schur */ 6286fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6287fa34dd3eSStefano Zampini if (zerodiag) { 6288fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 62894f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6290fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6291fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6292fa34dd3eSStefano Zampini } 6293fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6294fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6295fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6296fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6297fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6298fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6299fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6300fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6301fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6302fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6303fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6304fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6305fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6306fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6307fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6308fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6309fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6310fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6311fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6312fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6313fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6314fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6315fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6316fa34dd3eSStefano Zampini if (zerodiag) { 6317fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6318fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 63194f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6320fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6321fa34dd3eSStefano Zampini } 6322fa34dd3eSStefano Zampini /* BDDC */ 6323fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6324fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6325fa34dd3eSStefano Zampini 6326fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6327fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6328fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6329fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 63304f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 63314f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6332fa34dd3eSStefano Zampini } 63334f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6334fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6335fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6336fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6337fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6338fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6339fa34dd3eSStefano Zampini } 6340fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6341fa34dd3eSStefano Zampini } 6342