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 649*b3afcdbeSStefano Zampini /* no-net-flux */ 650*b3afcdbeSStefano Zampini if (pcbddc->benign_compute_nonetflux) { 651*b3afcdbeSStefano Zampini Mat_IS* matis = (Mat_IS*)(pc->pmat->data); 652*b3afcdbeSStefano Zampini MatNullSpace near_null_space; 653*b3afcdbeSStefano Zampini ISLocalToGlobalMapping rmap; 654*b3afcdbeSStefano Zampini Vec quad_vec; 655*b3afcdbeSStefano Zampini PetscScalar *pvals; 656*b3afcdbeSStefano Zampini PetscInt i,np,*dummyins; 657*b3afcdbeSStefano Zampini 658*b3afcdbeSStefano Zampini /* create vector to hold quadrature weights */ 659*b3afcdbeSStefano Zampini ierr = MatCreateVecs(pc->pmat,&quad_vec,NULL);CHKERRQ(ierr); 660*b3afcdbeSStefano Zampini ierr = VecSet(quad_vec,0.0);CHKERRQ(ierr); 661*b3afcdbeSStefano Zampini ierr = MatGetLocalToGlobalMapping(pc->pmat,&rmap,NULL);CHKERRQ(ierr); 662*b3afcdbeSStefano Zampini ierr = VecSetLocalToGlobalMapping(quad_vec,rmap);CHKERRQ(ierr); 663*b3afcdbeSStefano Zampini 664*b3afcdbeSStefano Zampini /* compute B^{(i)T} * 1_p */ 665*b3afcdbeSStefano Zampini np = 0; 666*b3afcdbeSStefano Zampini if (zerodiag) { 667*b3afcdbeSStefano Zampini ierr = ISGetLocalSize(zerodiag,&np);CHKERRQ(ierr); 668*b3afcdbeSStefano Zampini } else if (pressures) { 669*b3afcdbeSStefano Zampini ierr = ISGetLocalSize(pressures,&np);CHKERRQ(ierr); 670*b3afcdbeSStefano Zampini } 671*b3afcdbeSStefano Zampini ierr = PetscMalloc1(np,&pvals);CHKERRQ(ierr); 672*b3afcdbeSStefano Zampini for (i=0;i<np;i++) pvals[i] = 1.; 673*b3afcdbeSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 674*b3afcdbeSStefano Zampini if (np) { 675*b3afcdbeSStefano Zampini const PetscInt *pidxs; 676*b3afcdbeSStefano Zampini 677*b3afcdbeSStefano Zampini if (zerodiag) { 678*b3afcdbeSStefano Zampini ierr = ISGetIndices(zerodiag,&pidxs);CHKERRQ(ierr); 679*b3afcdbeSStefano Zampini } else if (pressures) { 680*b3afcdbeSStefano Zampini ierr = ISGetIndices(pressures,&pidxs);CHKERRQ(ierr); 681*b3afcdbeSStefano Zampini } 682*b3afcdbeSStefano Zampini ierr = VecSetValues(matis->x,np,pidxs,pvals,INSERT_VALUES);CHKERRQ(ierr); 683*b3afcdbeSStefano Zampini if (zerodiag) { 684*b3afcdbeSStefano Zampini ierr = ISRestoreIndices(zerodiag,&pidxs);CHKERRQ(ierr); 685*b3afcdbeSStefano Zampini } else if (pressures) { 686*b3afcdbeSStefano Zampini ierr = ISRestoreIndices(pressures,&pidxs);CHKERRQ(ierr); 687*b3afcdbeSStefano Zampini } 688*b3afcdbeSStefano Zampini } 689*b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 690*b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 691*b3afcdbeSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 692*b3afcdbeSStefano Zampini ierr = PetscFree(pvals);CHKERRQ(ierr); 693*b3afcdbeSStefano Zampini 694*b3afcdbeSStefano Zampini /* decide which of the sharing ranks (per dof) has to insert the values (should just be a matter of having a different orientation) */ 695*b3afcdbeSStefano Zampini ierr = MatISSetUpSF(pc->pmat);CHKERRQ(ierr); 696*b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = -1; 697*b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) matis->sf_leafdata[i] = PetscGlobalRank; 698*b3afcdbeSStefano Zampini ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 699*b3afcdbeSStefano Zampini ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_MAX);CHKERRQ(ierr); 700*b3afcdbeSStefano Zampini ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 701*b3afcdbeSStefano Zampini ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr); 702*b3afcdbeSStefano Zampini ierr = VecGetArray(matis->y,&pvals);CHKERRQ(ierr); 703*b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) { 704*b3afcdbeSStefano Zampini if (PetscGlobalRank != matis->sf_leafdata[i]) { 705*b3afcdbeSStefano Zampini pvals[i] = 0.; 706*b3afcdbeSStefano Zampini } 707*b3afcdbeSStefano Zampini } 708*b3afcdbeSStefano Zampini ierr = PetscMalloc1(matis->sf_nleaves,&dummyins);CHKERRQ(ierr); 709*b3afcdbeSStefano Zampini for (i=0;i<matis->sf_nleaves;i++) dummyins[i] = i; 710*b3afcdbeSStefano Zampini ierr = VecSetValuesLocal(quad_vec,matis->sf_nleaves,dummyins,pvals,ADD_VALUES);CHKERRQ(ierr); 711*b3afcdbeSStefano Zampini ierr = VecRestoreArray(matis->y,&pvals);CHKERRQ(ierr); 712*b3afcdbeSStefano Zampini ierr = PetscFree(dummyins);CHKERRQ(ierr); 713*b3afcdbeSStefano Zampini 714*b3afcdbeSStefano Zampini /* assembly quadrature vec and attach near null space to pmat */ 715*b3afcdbeSStefano Zampini ierr = VecAssemblyBegin(quad_vec);CHKERRQ(ierr); 716*b3afcdbeSStefano Zampini ierr = VecAssemblyEnd(quad_vec);CHKERRQ(ierr); 717*b3afcdbeSStefano Zampini ierr = VecNormalize(quad_vec,NULL);CHKERRQ(ierr); 718*b3afcdbeSStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)pc),PETSC_FALSE,1,&quad_vec,&near_null_space);CHKERRQ(ierr); 719*b3afcdbeSStefano Zampini ierr = VecDestroy(&quad_vec);CHKERRQ(ierr); 720*b3afcdbeSStefano Zampini ierr = MatSetNearNullSpace(pc->pmat,near_null_space);CHKERRQ(ierr); 721*b3afcdbeSStefano Zampini ierr = MatNullSpaceDestroy(&near_null_space);CHKERRQ(ierr); 722*b3afcdbeSStefano Zampini } 723*b3afcdbeSStefano Zampini 724*b3afcdbeSStefano Zampini /* change of basis and p0 dofs */ 7254f1b2e48SStefano Zampini if (has_null_pressures) { 7264f1b2e48SStefano Zampini IS zerodiagc; 7274f1b2e48SStefano Zampini const PetscInt *idxs,*idxsc; 7284f1b2e48SStefano Zampini PetscInt i,s,*nnz; 7294f1b2e48SStefano Zampini 7304f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag,&nz);CHKERRQ(ierr); 731339f8db1SStefano Zampini ierr = ISComplement(zerodiag,0,n,&zerodiagc);CHKERRQ(ierr); 732339f8db1SStefano Zampini ierr = ISGetIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 733339f8db1SStefano Zampini /* local change of basis for pressures */ 734339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 73597d764eeSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);CHKERRQ(ierr); 736339f8db1SStefano Zampini ierr = MatSetType(pcbddc->benign_change,MATAIJ);CHKERRQ(ierr); 737339f8db1SStefano Zampini ierr = MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 738339f8db1SStefano Zampini ierr = PetscMalloc1(n,&nnz);CHKERRQ(ierr); 7394f1b2e48SStefano Zampini for (i=0;i<n-nz;i++) nnz[idxsc[i]] = 1; /* identity on velocities plus pressure dofs for non-singular subdomains */ 7404f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 7414f1b2e48SStefano Zampini PetscInt nzs,j; 7424f1b2e48SStefano Zampini 7434f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[i],&nzs);CHKERRQ(ierr); 7444f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7454f1b2e48SStefano Zampini for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */ 7464f1b2e48SStefano Zampini nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */ 7474f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[i],&idxs);CHKERRQ(ierr); 7484f1b2e48SStefano Zampini } 749339f8db1SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);CHKERRQ(ierr); 750339f8db1SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 751339f8db1SStefano Zampini /* set identity on velocities */ 752339f8db1SStefano Zampini for (i=0;i<n-nz;i++) { 753339f8db1SStefano Zampini ierr = MatSetValue(pcbddc->benign_change,idxsc[i],idxsc[i],1.,INSERT_VALUES);CHKERRQ(ierr); 754339f8db1SStefano Zampini } 7554f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiagc,&idxsc);CHKERRQ(ierr); 7564f1b2e48SStefano Zampini ierr = ISDestroy(&zerodiagc);CHKERRQ(ierr); 7579f47a83aSStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 7584f1b2e48SStefano Zampini ierr = PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);CHKERRQ(ierr); 759339f8db1SStefano Zampini /* set change on pressures */ 7604f1b2e48SStefano Zampini for (s=0;s<pcbddc->benign_n;s++) { 7614f1b2e48SStefano Zampini PetscScalar *array; 7624f1b2e48SStefano Zampini PetscInt nzs; 7634f1b2e48SStefano Zampini 7644f1b2e48SStefano Zampini ierr = ISGetLocalSize(zerodiag_subs[s],&nzs);CHKERRQ(ierr); 7654f1b2e48SStefano Zampini ierr = ISGetIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 7664f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) { 767339f8db1SStefano Zampini PetscScalar vals[2]; 768339f8db1SStefano Zampini PetscInt cols[2]; 769339f8db1SStefano Zampini 770339f8db1SStefano Zampini cols[0] = idxs[i]; 7714f1b2e48SStefano Zampini cols[1] = idxs[nzs-1]; 772339f8db1SStefano Zampini vals[0] = 1.; 773b0f5fe93SStefano Zampini vals[1] = 1.; 7744f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 775339f8db1SStefano Zampini } 7764f1b2e48SStefano Zampini ierr = PetscMalloc1(nzs,&array);CHKERRQ(ierr); 7774f1b2e48SStefano Zampini for (i=0;i<nzs-1;i++) array[i] = -1.; 7784f1b2e48SStefano Zampini array[nzs-1] = 1.; 7794f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);CHKERRQ(ierr); 7804f1b2e48SStefano Zampini /* store local idxs for p0 */ 7814f1b2e48SStefano Zampini pcbddc->benign_p0_lidx[s] = idxs[nzs-1]; 7824f1b2e48SStefano Zampini ierr = ISRestoreIndices(zerodiag_subs[s],&idxs);CHKERRQ(ierr); 783339f8db1SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 7844f1b2e48SStefano Zampini } 785339f8db1SStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 786339f8db1SStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 787a3df083aSStefano Zampini /* project if needed */ 788a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 78997d764eeSStefano Zampini ierr = MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&pcbddc->benign_original_mat);CHKERRQ(ierr); 790339f8db1SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 79197d764eeSStefano Zampini ierr = MatSeqAIJCompress(pcbddc->benign_original_mat,&pcbddc->local_mat);CHKERRQ(ierr); 79297d764eeSStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 793a3df083aSStefano Zampini } 7944f1b2e48SStefano Zampini /* store global idxs for p0 */ 7954f1b2e48SStefano Zampini ierr = ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 796339f8db1SStefano Zampini } 797ca92afb2SStefano Zampini pcbddc->benign_zerodiag_subs = zerodiag_subs; 7984f1b2e48SStefano Zampini ierr = ISDestroy(&pressures);CHKERRQ(ierr); 799b0f5fe93SStefano Zampini 800b0f5fe93SStefano Zampini /* determines if the coarse solver will be singular or not */ 801b0f5fe93SStefano Zampini ierr = MPI_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 802339f8db1SStefano Zampini *zerodiaglocal = zerodiag; 803339f8db1SStefano Zampini PetscFunctionReturn(0); 804339f8db1SStefano Zampini } 805339f8db1SStefano Zampini 806339f8db1SStefano Zampini #undef __FUNCT__ 807015636ebSStefano Zampini #define __FUNCT__ "PCBDDCBenignGetOrSetP0" 808015636ebSStefano Zampini PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get) 809efc2fbd9SStefano Zampini { 810efc2fbd9SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 811efc2fbd9SStefano Zampini PetscErrorCode ierr; 812efc2fbd9SStefano Zampini 813efc2fbd9SStefano Zampini PetscFunctionBegin; 814efc2fbd9SStefano Zampini if (!pcbddc->benign_sf) { 815efc2fbd9SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);CHKERRQ(ierr); 8164f1b2e48SStefano Zampini ierr = PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);CHKERRQ(ierr); 817efc2fbd9SStefano Zampini } 818015636ebSStefano Zampini if (get) { /* use SF to get values */ 819efc2fbd9SStefano Zampini PetscScalar *array; 820efc2fbd9SStefano Zampini 821efc2fbd9SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8224f1b2e48SStefano Zampini ierr = PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 8234f1b2e48SStefano Zampini ierr = PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);CHKERRQ(ierr); 824efc2fbd9SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8254f1b2e48SStefano Zampini } else { /* use VecSetValues (not scalable, I should try to find a better solution (defining a new MPI_OP for reduction) */ 8264f1b2e48SStefano Zampini ierr = VecSetValues(v,pcbddc->benign_n,pcbddc->benign_p0_gidx,pcbddc->benign_p0,INSERT_VALUES);CHKERRQ(ierr); 827efc2fbd9SStefano Zampini ierr = VecAssemblyBegin(v);CHKERRQ(ierr); 828efc2fbd9SStefano Zampini ierr = VecAssemblyEnd(v);CHKERRQ(ierr); 829efc2fbd9SStefano Zampini } 830efc2fbd9SStefano Zampini PetscFunctionReturn(0); 831efc2fbd9SStefano Zampini } 832efc2fbd9SStefano Zampini 833efc2fbd9SStefano Zampini #undef __FUNCT__ 834c263805aSStefano Zampini #define __FUNCT__ "PCBDDCBenignPopOrPushB0" 835c263805aSStefano Zampini PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop) 836c263805aSStefano Zampini { 837c263805aSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 838c263805aSStefano Zampini PetscErrorCode ierr; 839c263805aSStefano Zampini 840c263805aSStefano Zampini PetscFunctionBegin; 841c263805aSStefano Zampini /* TODO: add error checking 842c263805aSStefano Zampini - avoid nested pop (or push) calls. 843c263805aSStefano Zampini - cannot push before pop. 8441c604dc7SStefano Zampini - cannot call this if pcbddc->local_mat is NULL 845c263805aSStefano Zampini */ 8464f1b2e48SStefano Zampini if (!pcbddc->benign_n) { 847efc2fbd9SStefano Zampini PetscFunctionReturn(0); 848efc2fbd9SStefano Zampini } 849c263805aSStefano Zampini if (pop) { 850a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 8514f1b2e48SStefano Zampini IS is_p0; 8524f1b2e48SStefano Zampini MatReuse reuse; 853c263805aSStefano Zampini 854c263805aSStefano Zampini /* extract B_0 */ 8554f1b2e48SStefano Zampini reuse = MAT_INITIAL_MATRIX; 8564f1b2e48SStefano Zampini if (pcbddc->benign_B0) { 8574f1b2e48SStefano Zampini reuse = MAT_REUSE_MATRIX; 8584f1b2e48SStefano Zampini } 8594f1b2e48SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 8604f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);CHKERRQ(ierr); 861c263805aSStefano Zampini /* remove rows and cols from local problem */ 862c263805aSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr); 86397d764eeSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 8644f1b2e48SStefano Zampini ierr = MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);CHKERRQ(ierr); 8654f1b2e48SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 866a3df083aSStefano Zampini } else { 867a3df083aSStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 868a3df083aSStefano Zampini PetscScalar *vals; 869a3df083aSStefano Zampini PetscInt i,n,*idxs_ins; 870a3df083aSStefano Zampini 871a3df083aSStefano Zampini ierr = VecGetLocalSize(matis->y,&n);CHKERRQ(ierr); 872a3df083aSStefano Zampini ierr = PetscMalloc2(n,&idxs_ins,n,&vals);CHKERRQ(ierr); 873a3df083aSStefano Zampini if (!pcbddc->benign_B0) { 8740b5adadeSStefano Zampini PetscInt *nnz; 875a3df083aSStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);CHKERRQ(ierr); 876a3df083aSStefano Zampini ierr = MatSetType(pcbddc->benign_B0,MATAIJ);CHKERRQ(ierr); 877a3df083aSStefano Zampini ierr = MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 878331e053bSStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&nnz);CHKERRQ(ierr); 879331e053bSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 880331e053bSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);CHKERRQ(ierr); 881331e053bSStefano Zampini nnz[i] = n - nnz[i]; 882331e053bSStefano Zampini } 883331e053bSStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);CHKERRQ(ierr); 884331e053bSStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 885331e053bSStefano Zampini } 886a3df083aSStefano Zampini 887a3df083aSStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 888a3df083aSStefano Zampini PetscScalar *array; 889a3df083aSStefano Zampini PetscInt *idxs,j,nz,cum; 890a3df083aSStefano Zampini 891a3df083aSStefano Zampini ierr = VecSet(matis->x,0.);CHKERRQ(ierr); 892a3df083aSStefano Zampini ierr = ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 893a3df083aSStefano Zampini ierr = ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 894a3df083aSStefano Zampini for (j=0;j<nz;j++) vals[j] = 1.; 895a3df083aSStefano Zampini ierr = VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 896a3df083aSStefano Zampini ierr = VecAssemblyBegin(matis->x);CHKERRQ(ierr); 897a3df083aSStefano Zampini ierr = VecAssemblyEnd(matis->x);CHKERRQ(ierr); 898a3df083aSStefano Zampini ierr = VecSet(matis->y,0.);CHKERRQ(ierr); 899a3df083aSStefano Zampini ierr = MatMult(matis->A,matis->x,matis->y);CHKERRQ(ierr); 900a3df083aSStefano Zampini ierr = VecGetArray(matis->y,&array);CHKERRQ(ierr); 901a3df083aSStefano Zampini cum = 0; 902a3df083aSStefano Zampini for (j=0;j<n;j++) { 903a3df083aSStefano Zampini if (PetscUnlikely(PetscAbsReal(array[j]) > PETSC_SMALL)) { 904a3df083aSStefano Zampini vals[cum] = array[j]; 905a3df083aSStefano Zampini idxs_ins[cum] = j; 906a3df083aSStefano Zampini cum++; 907a3df083aSStefano Zampini } 908a3df083aSStefano Zampini } 909a3df083aSStefano Zampini ierr = MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);CHKERRQ(ierr); 910a3df083aSStefano Zampini ierr = VecRestoreArray(matis->y,&array);CHKERRQ(ierr); 911a3df083aSStefano Zampini ierr = ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);CHKERRQ(ierr); 912a3df083aSStefano Zampini } 913a3df083aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 914a3df083aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 915a3df083aSStefano Zampini ierr = PetscFree2(idxs_ins,vals);CHKERRQ(ierr); 916a3df083aSStefano Zampini } 917c263805aSStefano Zampini } else { /* push */ 918a3df083aSStefano Zampini if (pcbddc->benign_change_explicit) { 9194f1b2e48SStefano Zampini PetscInt i; 9204f1b2e48SStefano Zampini 9214f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 9224f1b2e48SStefano Zampini PetscScalar *B0_vals; 9234f1b2e48SStefano Zampini PetscInt *B0_cols,B0_ncol; 9244f1b2e48SStefano Zampini 9254f1b2e48SStefano Zampini ierr = MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9264f1b2e48SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9277b034428SStefano Zampini ierr = MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);CHKERRQ(ierr); 9284f1b2e48SStefano Zampini ierr = MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);CHKERRQ(ierr); 9294f1b2e48SStefano Zampini ierr = MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);CHKERRQ(ierr); 9304f1b2e48SStefano Zampini } 931c263805aSStefano Zampini ierr = MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 932c263805aSStefano Zampini ierr = MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 933a3df083aSStefano Zampini } else { 934a3df083aSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!\n"); 935a3df083aSStefano Zampini } 936c263805aSStefano Zampini } 937c263805aSStefano Zampini PetscFunctionReturn(0); 938c263805aSStefano Zampini } 939c263805aSStefano Zampini 940c263805aSStefano Zampini #undef __FUNCT__ 941b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCAdaptiveSelection" 94208122e43SStefano Zampini PetscErrorCode PCBDDCAdaptiveSelection(PC pc) 943b1b3d7a2SStefano Zampini { 944b1b3d7a2SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 94508122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 94608122e43SStefano Zampini PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork; 94708122e43SStefano Zampini PetscBLASInt *B_iwork,*B_ifail; 94808122e43SStefano Zampini PetscScalar *work,lwork; 94908122e43SStefano Zampini PetscScalar *St,*S,*eigv; 95008122e43SStefano Zampini PetscScalar *Sarray,*Starray; 95108122e43SStefano Zampini PetscReal *eigs,thresh; 9521b968477SStefano Zampini PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs; 953f6f667cfSStefano Zampini PetscBool allocated_S_St; 95408122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 95508122e43SStefano Zampini PetscReal *rwork; 95608122e43SStefano Zampini #endif 957b1b3d7a2SStefano Zampini PetscErrorCode ierr; 958b1b3d7a2SStefano Zampini 959b1b3d7a2SStefano Zampini PetscFunctionBegin; 960df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 961df4d28bfSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO"); 96208122e43SStefano Zampini } 96308122e43SStefano Zampini 96406a4e24aSStefano Zampini if (sub_schurs->n_subs && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 96506a4e24aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for general matrix pencils (herm %d, posdef %d)\n",sub_schurs->is_hermitian,sub_schurs->is_posdef); 96606a4e24aSStefano Zampini } 96706a4e24aSStefano Zampini 968fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 969fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 970fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 971fd14bc51SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");CHKERRQ(ierr); 9721575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 973fd14bc51SStefano Zampini } 974fd14bc51SStefano Zampini 975e496cd5dSStefano Zampini if (pcbddc->dbg_flag) { 976e496cd5dSStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %d (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef); 977e496cd5dSStefano Zampini } 978e496cd5dSStefano Zampini 97908122e43SStefano Zampini /* max size of subsets */ 98008122e43SStefano Zampini mss = 0; 98108122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 98208122e43SStefano Zampini PetscInt subset_size; 983862806e4SStefano Zampini 98408122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 98508122e43SStefano Zampini mss = PetscMax(mss,subset_size); 98608122e43SStefano Zampini } 98708122e43SStefano Zampini 98808122e43SStefano Zampini /* min/max and threshold */ 98908122e43SStefano Zampini nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss; 990f6f667cfSStefano Zampini nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0; 99108122e43SStefano Zampini nmax = PetscMax(nmin,nmax); 992f6f667cfSStefano Zampini allocated_S_St = PETSC_FALSE; 993f6f667cfSStefano Zampini if (nmin) { 994f6f667cfSStefano Zampini allocated_S_St = PETSC_TRUE; 995f6f667cfSStefano Zampini } 99608122e43SStefano Zampini 99708122e43SStefano Zampini /* allocate lapack workspace */ 99808122e43SStefano Zampini cum = cum2 = 0; 99908122e43SStefano Zampini maxneigs = 0; 100008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 100108122e43SStefano Zampini PetscInt n,subset_size; 1002f6f667cfSStefano Zampini 100308122e43SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 100408122e43SStefano Zampini n = PetscMin(subset_size,nmax); 10059162d606SStefano Zampini cum += subset_size; 10069162d606SStefano Zampini cum2 += subset_size*n; 100708122e43SStefano Zampini maxneigs = PetscMax(maxneigs,n); 100808122e43SStefano Zampini } 100908122e43SStefano Zampini if (mss) { 10109ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 101108122e43SStefano Zampini PetscBLASInt B_itype = 1; 101208122e43SStefano Zampini PetscBLASInt B_N = mss; 10134c6709b3SStefano Zampini PetscReal zero = 0.0; 10144c6709b3SStefano Zampini PetscReal eps = 0.0; /* dlamch? */ 101508122e43SStefano Zampini 101608122e43SStefano Zampini B_lwork = -1; 101708122e43SStefano Zampini S = NULL; 101808122e43SStefano Zampini St = NULL; 1019a58a30b4SStefano Zampini eigs = NULL; 1020a58a30b4SStefano Zampini eigv = NULL; 1021a58a30b4SStefano Zampini B_iwork = NULL; 1022a58a30b4SStefano Zampini B_ifail = NULL; 1023d1710679SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1024d1710679SStefano Zampini rwork = NULL; 1025d1710679SStefano Zampini #endif 10268bec7fa6SStefano Zampini thresh = 1.0; 102708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 102808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 102908122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 103008122e43SStefano Zampini #else 103108122e43SStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr)); 103208122e43SStefano Zampini #endif 103308122e43SStefano Zampini if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr); 103408122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 103508122e43SStefano Zampini } else { 103608122e43SStefano Zampini /* TODO */ 103708122e43SStefano Zampini } 103808122e43SStefano Zampini } else { 103908122e43SStefano Zampini lwork = 0; 104008122e43SStefano Zampini } 104108122e43SStefano Zampini 104208122e43SStefano Zampini nv = 0; 1043d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */ 1044d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 104508122e43SStefano Zampini } 10464c6709b3SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 1047f6f667cfSStefano Zampini if (allocated_S_St) { 1048f6f667cfSStefano Zampini ierr = PetscMalloc2(mss*mss,&S,mss*mss,&St);CHKERRQ(ierr); 1049f6f667cfSStefano Zampini } 1050f6f667cfSStefano Zampini ierr = PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);CHKERRQ(ierr); 105108122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 105208122e43SStefano Zampini ierr = PetscMalloc1(7*mss,&rwork);CHKERRQ(ierr); 105308122e43SStefano Zampini #endif 10549162d606SStefano Zampini ierr = PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n, 10559162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr, 10569162d606SStefano Zampini nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr, 105708122e43SStefano Zampini nv+cum,&pcbddc->adaptive_constraints_idxs, 10589162d606SStefano Zampini nv+cum2,&pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 105908122e43SStefano Zampini ierr = PetscMemzero(pcbddc->adaptive_constraints_n,(nv+sub_schurs->n_subs)*sizeof(PetscInt));CHKERRQ(ierr); 106008122e43SStefano Zampini 106108122e43SStefano Zampini maxneigs = 0; 106272b8c272SStefano Zampini cum = cumarray = 0; 10639162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[0] = 0; 10649162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[0] = 0; 1065d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 106608122e43SStefano Zampini const PetscInt *idxs; 106708122e43SStefano Zampini 1068d62866d3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 106908122e43SStefano Zampini for (cum=0;cum<nv;cum++) { 107008122e43SStefano Zampini pcbddc->adaptive_constraints_n[cum] = 1; 107108122e43SStefano Zampini pcbddc->adaptive_constraints_idxs[cum] = idxs[cum]; 107208122e43SStefano Zampini pcbddc->adaptive_constraints_data[cum] = 1.0; 10739162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1; 10749162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1; 107508122e43SStefano Zampini } 1076d62866d3SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 107708122e43SStefano Zampini } 107808122e43SStefano Zampini 107908122e43SStefano Zampini if (mss) { /* multilevel */ 108008122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 108108122e43SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 108208122e43SStefano Zampini } 108308122e43SStefano Zampini 1084ffd830a3SStefano Zampini thresh = pcbddc->adaptive_threshold; 108508122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 108608122e43SStefano Zampini const PetscInt *idxs; 1087f6f667cfSStefano Zampini PetscReal infty = PETSC_MAX_REAL; 1088862806e4SStefano Zampini PetscInt j,subset_size,eigs_start = 0; 108908122e43SStefano Zampini PetscBLASInt B_N; 1090aff50787SStefano Zampini PetscBool same_data = PETSC_FALSE; 109108122e43SStefano Zampini 1092862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1093ffd830a3SStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 1094f6f667cfSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 1095f6f667cfSStefano Zampini if (allocated_S_St) { /* S and S_t should be copied since we could need them later */ 10969ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 1097aff50787SStefano Zampini PetscInt j,k; 1098aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscMemcmp later */ 1099aff50787SStefano Zampini ierr = PetscMemzero(S,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1100aff50787SStefano Zampini ierr = PetscMemzero(St,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 110108122e43SStefano Zampini } 110208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 1103aff50787SStefano Zampini for (k=j;k<subset_size;k++) { 1104aff50787SStefano Zampini S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k]; 1105aff50787SStefano Zampini St[j*subset_size+k] = Starray[cumarray+j*subset_size+k]; 1106aff50787SStefano Zampini } 110708122e43SStefano Zampini } 110808122e43SStefano Zampini } else { 110908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 111108122e43SStefano Zampini } 11128bec7fa6SStefano Zampini } else { 1113f6f667cfSStefano Zampini S = Sarray + cumarray; 1114f6f667cfSStefano Zampini St = Starray + cumarray; 11158bec7fa6SStefano Zampini } 1116aff50787SStefano Zampini /* see if we can save some work */ 1117aff50787SStefano Zampini if (sub_schurs->n_subs == 1) { 1118aff50787SStefano Zampini ierr = PetscMemcmp(S,St,subset_size*subset_size*sizeof(PetscScalar),&same_data);CHKERRQ(ierr); 1119aff50787SStefano Zampini } 1120aff50787SStefano Zampini 1121aff50787SStefano Zampini if (same_data) { /* there's no need of constraints here, deluxe scaling is enough */ 1122aff50787SStefano Zampini B_neigs = 0; 1123aff50787SStefano Zampini } else { 11249ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 112508122e43SStefano Zampini PetscBLASInt B_itype = 1; 1126f6f667cfSStefano Zampini PetscBLASInt B_IL, B_IU; 11274c6709b3SStefano Zampini PetscReal eps = -1.0; /* dlamch? */ 11289552c7c7SStefano Zampini PetscInt nmin_s; 112908122e43SStefano Zampini 1130fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 11318bec7fa6SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %d/%d %d %d.\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]); 1132fd14bc51SStefano Zampini } 1133d16cbb6bSStefano Zampini 113408122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1135d16cbb6bSStefano Zampini if (thresh > 1.+PETSC_SMALL) { 1136d16cbb6bSStefano Zampini 1137d16cbb6bSStefano Zampini /* ask for eigenvalues larger than thresh */ 113808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1139f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 114008122e43SStefano Zampini #else 1141f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 114208122e43SStefano Zampini #endif 1143d16cbb6bSStefano Zampini } else { 1144d16cbb6bSStefano Zampini B_IU = PetscMax(1,PetscMin(B_N,nmax)); 1145d16cbb6bSStefano Zampini B_IL = 1; 1146d16cbb6bSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1147d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 1148d16cbb6bSStefano Zampini #else 1149d16cbb6bSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 1150d16cbb6bSStefano Zampini #endif 1151d16cbb6bSStefano Zampini } 115208122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 115308122e43SStefano Zampini if (B_ierr) { 115408122e43SStefano Zampini if (B_ierr < 0 ) { 115508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 115608122e43SStefano Zampini } else if (B_ierr <= B_N) { 115708122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 115808122e43SStefano Zampini } else { 11599552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 116008122e43SStefano Zampini } 116108122e43SStefano Zampini } 116208122e43SStefano Zampini 116308122e43SStefano Zampini if (B_neigs > nmax) { 1164fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1165fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %d.\n",B_neigs,nmax); 1166fd14bc51SStefano Zampini } 1167f6f667cfSStefano Zampini eigs_start = B_neigs -nmax; 116808122e43SStefano Zampini B_neigs = nmax; 116908122e43SStefano Zampini } 117008122e43SStefano Zampini 11719552c7c7SStefano Zampini nmin_s = PetscMin(nmin,B_N); 11729552c7c7SStefano Zampini if (B_neigs < nmin_s) { 117308122e43SStefano Zampini PetscBLASInt B_neigs2; 117408122e43SStefano Zampini 1175f6f667cfSStefano Zampini B_IU = B_N - B_neigs; 1176f6f667cfSStefano Zampini B_IL = B_N - nmin_s + 1; 1177fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1178fd14bc51SStefano Zampini PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %d. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU); 1179fd14bc51SStefano Zampini } 11809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian) { 118108122e43SStefano Zampini PetscInt j; 118208122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118308122e43SStefano Zampini ierr = PetscMemcpy(S+j*(subset_size+1),Sarray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118408122e43SStefano Zampini } 118508122e43SStefano Zampini for (j=0;j<subset_size;j++) { 118608122e43SStefano Zampini ierr = PetscMemcpy(St+j*(subset_size+1),Starray+cumarray+j*(subset_size+1),(subset_size-j)*sizeof(PetscScalar));CHKERRQ(ierr); 118708122e43SStefano Zampini } 118808122e43SStefano Zampini } else { 118908122e43SStefano Zampini ierr = PetscMemcpy(S,Sarray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119008122e43SStefano Zampini ierr = PetscMemcpy(St,Starray+cumarray,subset_size*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 119108122e43SStefano Zampini } 119208122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 119308122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 1194f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr)); 119508122e43SStefano Zampini #else 1196f6f667cfSStefano Zampini PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&thresh,&infty,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr)); 119708122e43SStefano Zampini #endif 119808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 119908122e43SStefano Zampini B_neigs += B_neigs2; 120008122e43SStefano Zampini } 120108122e43SStefano Zampini if (B_ierr) { 120208122e43SStefano Zampini if (B_ierr < 0 ) { 120308122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr); 120408122e43SStefano Zampini } else if (B_ierr <= B_N) { 120508122e43SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr); 120608122e43SStefano Zampini } else { 12079552c7c7SStefano Zampini SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1); 120808122e43SStefano Zampini } 120908122e43SStefano Zampini } 1210fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1211ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);CHKERRQ(ierr); 121208122e43SStefano Zampini for (j=0;j<B_neigs;j++) { 121308122e43SStefano Zampini if (eigs[j] == 0.0) { 1214ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");CHKERRQ(ierr); 121508122e43SStefano Zampini } else { 1216ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);CHKERRQ(ierr); 1217fd14bc51SStefano Zampini } 121808122e43SStefano Zampini } 121908122e43SStefano Zampini } 122008122e43SStefano Zampini } else { 122108122e43SStefano Zampini /* TODO */ 122208122e43SStefano Zampini } 1223aff50787SStefano Zampini } 12246c3e6151SStefano Zampini /* change the basis back to the original one */ 12256c3e6151SStefano Zampini if (sub_schurs->change) { 122672b8c272SStefano Zampini Mat change,phi,phit; 12276c3e6151SStefano Zampini 12286c3e6151SStefano Zampini if (pcbddc->dbg_flag > 1) { 12296c3e6151SStefano Zampini PetscInt ii; 12306c3e6151SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 12316c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12326c3e6151SStefano Zampini for (j=0;j<B_N;j++) { 12336c3e6151SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);CHKERRQ(ierr); 12346c3e6151SStefano Zampini } 12356c3e6151SStefano Zampini } 12366c3e6151SStefano Zampini } 123772b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change,NULL);CHKERRQ(ierr); 12386c3e6151SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);CHKERRQ(ierr); 123972b8c272SStefano Zampini ierr = MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);CHKERRQ(ierr); 12406c3e6151SStefano Zampini ierr = MatCopy(phi,phit,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 12416c3e6151SStefano Zampini ierr = MatDestroy(&phit);CHKERRQ(ierr); 12426c3e6151SStefano Zampini ierr = MatDestroy(&phi);CHKERRQ(ierr); 12436c3e6151SStefano Zampini } 12448bec7fa6SStefano Zampini maxneigs = PetscMax(B_neigs,maxneigs); 12458bec7fa6SStefano Zampini pcbddc->adaptive_constraints_n[i+nv] = B_neigs; 12469162d606SStefano Zampini if (B_neigs) { 12479162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size*sizeof(PetscScalar));CHKERRQ(ierr); 1248fd14bc51SStefano Zampini 1249fd14bc51SStefano Zampini if (pcbddc->dbg_flag > 1) { 12509552c7c7SStefano Zampini PetscInt ii; 12519552c7c7SStefano Zampini for (ii=0;ii<B_neigs;ii++) { 1252ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);CHKERRQ(ierr); 12539552c7c7SStefano Zampini for (j=0;j<B_N;j++) { 1254ac47001eSStefano Zampini #if defined(PETSC_USE_COMPLEX) 1255ac47001eSStefano Zampini PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1256ac47001eSStefano Zampini PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]); 1257ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);CHKERRQ(ierr); 1258ac47001eSStefano Zampini #else 1259ac47001eSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);CHKERRQ(ierr); 1260ac47001eSStefano Zampini #endif 12619552c7c7SStefano Zampini } 12629552c7c7SStefano Zampini } 1263fd14bc51SStefano Zampini } 12649162d606SStefano Zampini ierr = PetscMemcpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 12659162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size; 12669162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs; 12679162d606SStefano Zampini cum++; 126808122e43SStefano Zampini } 126908122e43SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 127008122e43SStefano Zampini /* shift for next computation */ 127108122e43SStefano Zampini cumarray += subset_size*subset_size; 127208122e43SStefano Zampini } 1273fd14bc51SStefano Zampini if (pcbddc->dbg_flag) { 1274fd14bc51SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 1275fd14bc51SStefano Zampini } 127608122e43SStefano Zampini 127708122e43SStefano Zampini if (mss) { 127808122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);CHKERRQ(ierr); 127908122e43SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);CHKERRQ(ierr); 1280f6f667cfSStefano Zampini /* destroy matrices (junk) */ 1281f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1282f6f667cfSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 128308122e43SStefano Zampini } 1284f6f667cfSStefano Zampini if (allocated_S_St) { 1285f6f667cfSStefano Zampini ierr = PetscFree2(S,St);CHKERRQ(ierr); 1286f6f667cfSStefano Zampini } 1287f6f667cfSStefano Zampini ierr = PetscFree5(eigv,eigs,work,B_iwork,B_ifail);CHKERRQ(ierr); 128808122e43SStefano Zampini #if defined(PETSC_USE_COMPLEX) 128908122e43SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 129008122e43SStefano Zampini #endif 129108122e43SStefano Zampini if (pcbddc->dbg_flag) { 12921b968477SStefano Zampini PetscInt maxneigs_r; 129308122e43SStefano Zampini ierr = MPI_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 12949b28b941SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %d\n",maxneigs_r);CHKERRQ(ierr); 129508122e43SStefano Zampini } 129608122e43SStefano Zampini PetscFunctionReturn(0); 129708122e43SStefano Zampini } 1298b1b3d7a2SStefano Zampini 1299674ae819SStefano Zampini #undef __FUNCT__ 1300c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSolvers" 1301c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpSolvers(PC pc) 1302c8587f34SStefano Zampini { 1303c8587f34SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 13048629588bSStefano Zampini PetscScalar *coarse_submat_vals; 1305c8587f34SStefano Zampini PetscErrorCode ierr; 1306c8587f34SStefano Zampini 1307c8587f34SStefano Zampini PetscFunctionBegin; 1308f4ddd8eeSStefano Zampini /* Setup local scatters R_to_B and (optionally) R_to_D */ 13095e8657edSStefano Zampini /* PCBDDCSetUpLocalWorkVectors should be called first! */ 1310c8587f34SStefano Zampini ierr = PCBDDCSetUpLocalScatters(pc);CHKERRQ(ierr); 1311c8587f34SStefano Zampini 1312684f6988SStefano Zampini /* Setup local neumann solver ksp_R */ 13130fccc4e9SStefano Zampini /* PCBDDCSetUpLocalScatters should be called first! */ 1314684f6988SStefano Zampini ierr = PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 1315c8587f34SStefano Zampini 1316c8587f34SStefano Zampini /* Change global null space passed in by the user if change of basis has been requested */ 1317b9b85e73SStefano Zampini if (pcbddc->NullSpace && pcbddc->ChangeOfBasisMatrix) { 1318c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAdaptGlobal(pc);CHKERRQ(ierr); 1319c8587f34SStefano Zampini } 1320c8587f34SStefano Zampini 13218629588bSStefano Zampini /* 13228629588bSStefano Zampini Setup local correction and local part of coarse basis. 13238629588bSStefano Zampini Gives back the dense local part of the coarse matrix in column major ordering 13248629588bSStefano Zampini */ 132547f4ddc3SStefano Zampini ierr = PCBDDCSetUpCorrection(pc,&coarse_submat_vals);CHKERRQ(ierr); 13268629588bSStefano Zampini 13278629588bSStefano Zampini /* Compute total number of coarse nodes and setup coarse solver */ 13288629588bSStefano Zampini ierr = PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);CHKERRQ(ierr); 13298629588bSStefano Zampini 13308629588bSStefano Zampini /* free */ 13318629588bSStefano Zampini ierr = PetscFree(coarse_submat_vals);CHKERRQ(ierr); 1332c8587f34SStefano Zampini PetscFunctionReturn(0); 1333c8587f34SStefano Zampini } 1334c8587f34SStefano Zampini 1335c8587f34SStefano Zampini #undef __FUNCT__ 1336674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetCustomization" 1337674ae819SStefano Zampini PetscErrorCode PCBDDCResetCustomization(PC pc) 1338674ae819SStefano Zampini { 1339674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1340674ae819SStefano Zampini PetscErrorCode ierr; 1341674ae819SStefano Zampini 1342674ae819SStefano Zampini PetscFunctionBegin; 1343674ae819SStefano Zampini ierr = PCBDDCGraphResetCSR(pcbddc->mat_graph);CHKERRQ(ierr); 1344674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr); 134530368db7SStefano Zampini ierr = ISDestroy(&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 1346674ae819SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr); 1347674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr); 1348785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 1349674ae819SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr); 1350f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 1351f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 1352785d1243SStefano Zampini ierr = ISDestroy(&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 135363602bcaSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc,0,NULL);CHKERRQ(ierr); 135463602bcaSStefano Zampini ierr = PCBDDCSetDofsSplittingLocal(pc,0,NULL);CHKERRQ(ierr); 1355674ae819SStefano Zampini PetscFunctionReturn(0); 1356674ae819SStefano Zampini } 1357674ae819SStefano Zampini 1358674ae819SStefano Zampini #undef __FUNCT__ 1359674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetTopography" 1360674ae819SStefano Zampini PetscErrorCode PCBDDCResetTopography(PC pc) 1361674ae819SStefano Zampini { 1362674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 13634f1b2e48SStefano Zampini PetscInt i; 1364674ae819SStefano Zampini PetscErrorCode ierr; 1365674ae819SStefano Zampini 1366674ae819SStefano Zampini PetscFunctionBegin; 1367b9b85e73SStefano Zampini ierr = MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 1368674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 1369674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 1370674ae819SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 13714f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 13724f1b2e48SStefano Zampini ierr = ISDestroy(&pcbddc->local_subs[i]);CHKERRQ(ierr); 13734f1b2e48SStefano Zampini } 13744f1b2e48SStefano Zampini ierr = PetscFree(pcbddc->local_subs);CHKERRQ(ierr); 1375b96c3477SStefano Zampini ierr = PCBDDCSubSchursReset(pcbddc->sub_schurs);CHKERRQ(ierr); 1376674ae819SStefano Zampini PetscFunctionReturn(0); 1377674ae819SStefano Zampini } 1378674ae819SStefano Zampini 1379674ae819SStefano Zampini #undef __FUNCT__ 1380674ae819SStefano Zampini #define __FUNCT__ "PCBDDCResetSolvers" 1381674ae819SStefano Zampini PetscErrorCode PCBDDCResetSolvers(PC pc) 1382674ae819SStefano Zampini { 1383674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 1384674ae819SStefano Zampini PetscErrorCode ierr; 1385674ae819SStefano Zampini 1386674ae819SStefano Zampini PetscFunctionBegin; 1387674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 138858da7f69SStefano Zampini if (pcbddc->coarse_phi_B) { 1389ca92afb2SStefano Zampini PetscScalar *array; 139006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&array);CHKERRQ(ierr); 139106656605SStefano Zampini ierr = PetscFree(array);CHKERRQ(ierr); 139258da7f69SStefano Zampini } 1393674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1394674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 139515aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 139615aaf578SStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1397674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 1398674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 1399674ae819SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 140006656605SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 1401674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1402674ae819SStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14038ce42a96SStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 1404674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 1405674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 1406674ae819SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 1407f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr); 1408f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr); 1409f4ddd8eeSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 1410f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 1411727cdba6SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 14120e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 1413f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 141470cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 14156e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 141681d14e9dSStefano Zampini ierr = MatDestroy(&pcbddc->benign_change);CHKERRQ(ierr); 14170369aaf7SStefano Zampini ierr = VecDestroy(&pcbddc->benign_vec);CHKERRQ(ierr); 14188b9f24d4SStefano Zampini ierr = MatDestroy(&pcbddc->benign_original_mat);CHKERRQ(ierr); 14194f1b2e48SStefano Zampini ierr = MatDestroy(&pcbddc->benign_B0);CHKERRQ(ierr); 14208b9f24d4SStefano Zampini ierr = PetscSFDestroy(&pcbddc->benign_sf);CHKERRQ(ierr); 1421ca92afb2SStefano Zampini if (pcbddc->benign_zerodiag_subs) { 1422ca92afb2SStefano Zampini PetscInt i; 1423ca92afb2SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 1424ca92afb2SStefano Zampini ierr = ISDestroy(&pcbddc->benign_zerodiag_subs[i]);CHKERRQ(ierr); 1425ca92afb2SStefano Zampini } 1426ca92afb2SStefano Zampini ierr = PetscFree(pcbddc->benign_zerodiag_subs);CHKERRQ(ierr); 1427ca92afb2SStefano Zampini } 14284f1b2e48SStefano Zampini ierr = PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);CHKERRQ(ierr); 1429674ae819SStefano Zampini PetscFunctionReturn(0); 1430674ae819SStefano Zampini } 1431674ae819SStefano Zampini 1432674ae819SStefano Zampini #undef __FUNCT__ 1433f4ddd8eeSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalWorkVectors" 1434f4ddd8eeSStefano Zampini PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc) 14356bfb1811SStefano Zampini { 14366bfb1811SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 14376bfb1811SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 14386bfb1811SStefano Zampini VecType impVecType; 14394f1b2e48SStefano Zampini PetscInt n_constraints,n_R,old_size; 14406bfb1811SStefano Zampini PetscErrorCode ierr; 14416bfb1811SStefano Zampini 14426bfb1811SStefano Zampini PetscFunctionBegin; 1443f4ddd8eeSStefano Zampini if (!pcbddc->ConstraintMatrix) { 1444019a44ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC Constraint matrix has not been created"); 1445f4ddd8eeSStefano Zampini } 1446e7b262bdSStefano Zampini /* get sizes */ 14474f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices; 1448b371cd4fSStefano Zampini n_R = pcis->n - pcbddc->n_vertices; 14496bfb1811SStefano Zampini ierr = VecGetType(pcis->vec1_N,&impVecType);CHKERRQ(ierr); 1450e7b262bdSStefano Zampini /* local work vectors (try to avoid unneeded work)*/ 1451e7b262bdSStefano Zampini /* R nodes */ 1452e7b262bdSStefano Zampini old_size = -1; 1453e7b262bdSStefano Zampini if (pcbddc->vec1_R) { 1454e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_R,&old_size);CHKERRQ(ierr); 1455e7b262bdSStefano Zampini } 1456e7b262bdSStefano Zampini if (n_R != old_size) { 1457e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_R);CHKERRQ(ierr); 1458e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec2_R);CHKERRQ(ierr); 14596bfb1811SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);CHKERRQ(ierr); 14606bfb1811SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);CHKERRQ(ierr); 14616bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_R,impVecType);CHKERRQ(ierr); 14626bfb1811SStefano Zampini ierr = VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);CHKERRQ(ierr); 1463e7b262bdSStefano Zampini } 1464e7b262bdSStefano Zampini /* local primal dofs */ 1465e7b262bdSStefano Zampini old_size = -1; 1466e7b262bdSStefano Zampini if (pcbddc->vec1_P) { 1467e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_P,&old_size);CHKERRQ(ierr); 1468e7b262bdSStefano Zampini } 1469e9189074SStefano Zampini if (pcbddc->local_primal_size != old_size) { 1470e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_P);CHKERRQ(ierr); 147183b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);CHKERRQ(ierr); 1472e9189074SStefano Zampini ierr = VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);CHKERRQ(ierr); 14736bfb1811SStefano Zampini ierr = VecSetType(pcbddc->vec1_P,impVecType);CHKERRQ(ierr); 1474e7b262bdSStefano Zampini } 1475e7b262bdSStefano Zampini /* local explicit constraints */ 1476e7b262bdSStefano Zampini old_size = -1; 1477e7b262bdSStefano Zampini if (pcbddc->vec1_C) { 1478e7b262bdSStefano Zampini ierr = VecGetSize(pcbddc->vec1_C,&old_size);CHKERRQ(ierr); 1479e7b262bdSStefano Zampini } 1480e7b262bdSStefano Zampini if (n_constraints && n_constraints != old_size) { 1481e7b262bdSStefano Zampini ierr = VecDestroy(&pcbddc->vec1_C);CHKERRQ(ierr); 148283b7ccabSStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);CHKERRQ(ierr); 148383b7ccabSStefano Zampini ierr = VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);CHKERRQ(ierr); 148483b7ccabSStefano Zampini ierr = VecSetType(pcbddc->vec1_C,impVecType);CHKERRQ(ierr); 148583b7ccabSStefano Zampini } 14866bfb1811SStefano Zampini PetscFunctionReturn(0); 14876bfb1811SStefano Zampini } 14886bfb1811SStefano Zampini 14896bfb1811SStefano Zampini #undef __FUNCT__ 149047f4ddc3SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCorrection" 149147f4ddc3SStefano Zampini PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n) 149288ebb749SStefano Zampini { 149325084f0cSStefano Zampini PetscErrorCode ierr; 149425084f0cSStefano Zampini /* pointers to pcis and pcbddc */ 149588ebb749SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 149688ebb749SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 1497d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 149825084f0cSStefano Zampini /* submatrices of local problem */ 149980677318SStefano Zampini Mat A_RV,A_VR,A_VV,local_auxmat2_R; 150006656605SStefano Zampini /* submatrices of local coarse problem */ 150106656605SStefano Zampini Mat S_VV,S_CV,S_VC,S_CC; 150225084f0cSStefano Zampini /* working matrices */ 150306656605SStefano Zampini Mat C_CR; 150425084f0cSStefano Zampini /* additional working stuff */ 150506656605SStefano Zampini PC pc_R; 15064f1b2e48SStefano Zampini Mat F; 1507a3df083aSStefano Zampini PetscBool isLU,isCHOL,isILU,need_benign_correction; 150806656605SStefano Zampini 150925084f0cSStefano Zampini PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */ 151006656605SStefano Zampini PetscScalar *work; 151106656605SStefano Zampini PetscInt *idx_V_B; 1512ffd830a3SStefano Zampini PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I; 151306656605SStefano Zampini PetscInt i,n_R,n_D,n_B; 1514ffd830a3SStefano Zampini 151525084f0cSStefano Zampini /* some shortcuts to scalars */ 151606656605SStefano Zampini PetscScalar one=1.0,m_one=-1.0; 151788ebb749SStefano Zampini 151888ebb749SStefano Zampini PetscFunctionBegin; 1519ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal && pcbddc->benign_n) { 1520ffd830a3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented"); 1521ffd830a3SStefano Zampini } 1522ffd830a3SStefano Zampini 1523ffd830a3SStefano Zampini /* Set Non-overlapping dimensions */ 1524b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 15254f1b2e48SStefano Zampini n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices; 1526b371cd4fSStefano Zampini n_B = pcis->n_B; 1527b371cd4fSStefano Zampini n_D = pcis->n - n_B; 152888ebb749SStefano Zampini n_R = pcis->n - n_vertices; 152988ebb749SStefano Zampini 153088ebb749SStefano Zampini /* vertices in boundary numbering */ 1531785e854fSJed Brown ierr = PetscMalloc1(n_vertices,&idx_V_B);CHKERRQ(ierr); 15320e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);CHKERRQ(ierr); 153388ebb749SStefano Zampini if (i != n_vertices) { 153422d5777bSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %d != %d\n",n_vertices,i); 153588ebb749SStefano Zampini } 153688ebb749SStefano Zampini 153706656605SStefano Zampini /* Subdomain contribution (Non-overlapping) to coarse matrix */ 1538019a44ceSStefano Zampini ierr = PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);CHKERRQ(ierr); 153906656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);CHKERRQ(ierr); 154006656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);CHKERRQ(ierr); 154106656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);CHKERRQ(ierr); 154206656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);CHKERRQ(ierr); 154306656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);CHKERRQ(ierr); 154406656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);CHKERRQ(ierr); 154506656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);CHKERRQ(ierr); 154606656605SStefano Zampini ierr = MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);CHKERRQ(ierr); 154706656605SStefano Zampini 154806656605SStefano Zampini /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */ 154906656605SStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_R);CHKERRQ(ierr); 155006656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);CHKERRQ(ierr); 155106656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCILU,&isILU);CHKERRQ(ierr); 155206656605SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);CHKERRQ(ierr); 1553ffd830a3SStefano Zampini lda_rhs = n_R; 1554a3df083aSStefano Zampini need_benign_correction = PETSC_FALSE; 155506656605SStefano Zampini if (isLU || isILU || isCHOL) { 155606656605SStefano Zampini ierr = PCFactorGetMatrix(pc_R,&F);CHKERRQ(ierr); 1557df4d28bfSStefano Zampini } else if (sub_schurs->reuse_solver) { 1558df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1559d62866d3SStefano Zampini MatFactorType type; 1560d62866d3SStefano Zampini 1561df4d28bfSStefano Zampini F = reuse_solver->F; 15626816873aSStefano Zampini ierr = MatGetFactorType(F,&type);CHKERRQ(ierr); 1563d62866d3SStefano Zampini if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE; 1564ffd830a3SStefano Zampini ierr = MatGetSize(F,&lda_rhs,NULL);CHKERRQ(ierr); 1565df4d28bfSStefano Zampini need_benign_correction = !!reuse_solver->benign_n; 156606656605SStefano Zampini } else { 156706656605SStefano Zampini F = NULL; 156806656605SStefano Zampini } 156906656605SStefano Zampini 1570ffd830a3SStefano Zampini /* allocate workspace */ 1571ffd830a3SStefano Zampini n = 0; 1572ffd830a3SStefano Zampini if (n_constraints) { 1573ffd830a3SStefano Zampini n += lda_rhs*n_constraints; 1574ffd830a3SStefano Zampini } 1575ffd830a3SStefano Zampini if (n_vertices) { 1576ffd830a3SStefano Zampini n = PetscMax(2*lda_rhs*n_vertices,n); 1577ffd830a3SStefano Zampini n = PetscMax((lda_rhs+n_B)*n_vertices,n); 1578ffd830a3SStefano Zampini } 1579ffd830a3SStefano Zampini ierr = PetscMalloc1(n,&work);CHKERRQ(ierr); 1580ffd830a3SStefano Zampini 158188ebb749SStefano Zampini /* Precompute stuffs needed for preprocessing and application of BDDC*/ 158288ebb749SStefano Zampini if (n_constraints) { 158372b8c272SStefano Zampini Mat M1,M2,M3,C_B; 158406656605SStefano Zampini IS is_aux; 158580677318SStefano Zampini PetscScalar *array,*array2; 158606656605SStefano Zampini 1587f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat1);CHKERRQ(ierr); 158880677318SStefano Zampini ierr = MatDestroy(&pcbddc->local_auxmat2);CHKERRQ(ierr); 158988ebb749SStefano Zampini 159025084f0cSStefano Zampini /* Extract constraints on R nodes: C_{CR} */ 159125084f0cSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);CHKERRQ(ierr); 15928ce42a96SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);CHKERRQ(ierr); 159372b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);CHKERRQ(ierr); 159488ebb749SStefano Zampini 159580677318SStefano Zampini /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */ 159680677318SStefano Zampini /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC application */ 1597ffd830a3SStefano Zampini ierr = PetscMemzero(work,lda_rhs*n_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 159888ebb749SStefano Zampini for (i=0;i<n_constraints;i++) { 159906656605SStefano Zampini const PetscScalar *row_cmat_values; 160006656605SStefano Zampini const PetscInt *row_cmat_indices; 160106656605SStefano Zampini PetscInt size_of_constraint,j; 160288ebb749SStefano Zampini 160306656605SStefano Zampini ierr = MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 160406656605SStefano Zampini for (j=0;j<size_of_constraint;j++) { 1605ffd830a3SStefano Zampini work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j]; 160606656605SStefano Zampini } 160706656605SStefano Zampini ierr = MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);CHKERRQ(ierr); 160806656605SStefano Zampini } 1609ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);CHKERRQ(ierr); 161006656605SStefano Zampini if (F) { 161106656605SStefano Zampini Mat B; 161206656605SStefano Zampini 1613ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 1614a3df083aSStefano Zampini if (need_benign_correction) { 1615df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1616a3df083aSStefano Zampini 161772b8c272SStefano Zampini /* rhs is already zero on interior dofs, no need to change the rhs */ 161872b8c272SStefano Zampini ierr = PetscMemzero(reuse_solver->benign_save_vals,pcbddc->benign_n*sizeof(PetscScalar));CHKERRQ(ierr); 1619a3df083aSStefano Zampini } 162080677318SStefano Zampini ierr = MatMatSolve(F,B,local_auxmat2_R);CHKERRQ(ierr); 1621a3df083aSStefano Zampini if (need_benign_correction) { 1622a3df083aSStefano Zampini PetscScalar *marr; 1623df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1624a3df083aSStefano Zampini 1625a3df083aSStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1626a3df083aSStefano Zampini for (i=0;i<n_constraints;i++) { 1627a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1628df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1629a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1630a3df083aSStefano Zampini } 1631a3df083aSStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 1632a3df083aSStefano Zampini } 163306656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 163406656605SStefano Zampini } else { 163580677318SStefano Zampini PetscScalar *marr; 163680677318SStefano Zampini 163780677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 163806656605SStefano Zampini for (i=0;i<n_constraints;i++) { 1639ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1640ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);CHKERRQ(ierr); 164106656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 164206656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 164306656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 164406656605SStefano Zampini } 164580677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&marr);CHKERRQ(ierr); 164606656605SStefano Zampini } 164780677318SStefano Zampini if (!pcbddc->switch_static) { 164880677318SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);CHKERRQ(ierr); 164980677318SStefano Zampini ierr = MatDenseGetArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 165080677318SStefano Zampini ierr = MatDenseGetArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 165180677318SStefano Zampini for (i=0;i<n_constraints;i++) { 1652ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);CHKERRQ(ierr); 165380677318SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,array+i*n_B);CHKERRQ(ierr); 165480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 165680677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 165780677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 165880677318SStefano Zampini } 165980677318SStefano Zampini ierr = MatDenseRestoreArray(local_auxmat2_R,&array2);CHKERRQ(ierr); 166080677318SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->local_auxmat2,&array);CHKERRQ(ierr); 166172b8c272SStefano Zampini ierr = MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 166280677318SStefano Zampini } else { 1663ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1664ffd830a3SStefano Zampini IS dummy; 1665ffd830a3SStefano Zampini 1666ffd830a3SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);CHKERRQ(ierr); 166772b8c272SStefano Zampini ierr = MatGetSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);CHKERRQ(ierr); 1668ffd830a3SStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 1669ffd830a3SStefano Zampini } else { 167080677318SStefano Zampini ierr = PetscObjectReference((PetscObject)local_auxmat2_R);CHKERRQ(ierr); 167180677318SStefano Zampini pcbddc->local_auxmat2 = local_auxmat2_R; 1672ffd830a3SStefano Zampini } 167325084f0cSStefano Zampini ierr = MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);CHKERRQ(ierr); 167480677318SStefano Zampini } 167580677318SStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 167680677318SStefano Zampini /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */ 167780677318SStefano Zampini ierr = MatScale(M3,m_one);CHKERRQ(ierr); 167806656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M1);CHKERRQ(ierr); 167906656605SStefano Zampini ierr = MatDuplicate(M3,MAT_DO_NOT_COPY_VALUES,&M2);CHKERRQ(ierr); 168080677318SStefano Zampini if (isCHOL) { 168180677318SStefano Zampini ierr = MatCholeskyFactor(M3,NULL,NULL);CHKERRQ(ierr); 168280677318SStefano Zampini } else { 168325084f0cSStefano Zampini ierr = MatLUFactor(M3,NULL,NULL,NULL);CHKERRQ(ierr); 168480677318SStefano Zampini } 168580677318SStefano Zampini ierr = VecSet(pcbddc->vec1_C,one);CHKERRQ(ierr); 168606656605SStefano Zampini ierr = MatDiagonalSet(M2,pcbddc->vec1_C,INSERT_VALUES);CHKERRQ(ierr); 168725084f0cSStefano Zampini ierr = MatMatSolve(M3,M2,M1);CHKERRQ(ierr); 168825084f0cSStefano Zampini ierr = MatDestroy(&M2);CHKERRQ(ierr); 168925084f0cSStefano Zampini ierr = MatDestroy(&M3);CHKERRQ(ierr); 169080677318SStefano Zampini /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC application in KSP and in preproc */ 169172b8c272SStefano Zampini ierr = MatMatMult(M1,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);CHKERRQ(ierr); 169272b8c272SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 169306656605SStefano Zampini ierr = MatCopy(M1,S_CC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); /* S_CC can have a different LDA, MatMatSolve doesn't support it */ 169406656605SStefano Zampini ierr = MatDestroy(&M1);CHKERRQ(ierr); 1695f4ddd8eeSStefano Zampini } 1696fc227af8SStefano Zampini 1697fc227af8SStefano Zampini /* Get submatrices from subdomain matrix */ 169888ebb749SStefano Zampini if (n_vertices) { 169906656605SStefano Zampini IS is_aux; 17003a50541eSStefano Zampini 1701df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */ 17026816873aSStefano Zampini IS tis; 17036816873aSStefano Zampini 17046816873aSStefano Zampini ierr = ISDuplicate(pcbddc->is_R_local,&tis);CHKERRQ(ierr); 17056816873aSStefano Zampini ierr = ISSort(tis);CHKERRQ(ierr); 17066816873aSStefano Zampini ierr = ISComplement(tis,0,pcis->n,&is_aux);CHKERRQ(ierr); 17076816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 17086816873aSStefano Zampini } else { 17093a50541eSStefano Zampini ierr = ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);CHKERRQ(ierr); 17106816873aSStefano Zampini } 17119577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);CHKERRQ(ierr); 17129577ea80SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);CHKERRQ(ierr); 171304708bb6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);CHKERRQ(ierr); 171425084f0cSStefano Zampini ierr = ISDestroy(&is_aux);CHKERRQ(ierr); 171588ebb749SStefano Zampini } 171688ebb749SStefano Zampini 171788ebb749SStefano Zampini /* Matrix of coarse basis functions (local) */ 1718f4ddd8eeSStefano Zampini if (pcbddc->coarse_phi_B) { 171906656605SStefano Zampini PetscInt on_B,on_primal,on_D=n_D; 172006656605SStefano Zampini if (pcbddc->coarse_phi_D) { 172106656605SStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);CHKERRQ(ierr); 172206656605SStefano Zampini } 1723f4ddd8eeSStefano Zampini ierr = MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);CHKERRQ(ierr); 172406656605SStefano Zampini if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) { 172506656605SStefano Zampini PetscScalar *marray; 172606656605SStefano Zampini 172706656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&marray);CHKERRQ(ierr); 172806656605SStefano Zampini ierr = PetscFree(marray);CHKERRQ(ierr); 1729f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_B);CHKERRQ(ierr); 1730f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_B);CHKERRQ(ierr); 1731f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_phi_D);CHKERRQ(ierr); 1732f4ddd8eeSStefano Zampini ierr = MatDestroy(&pcbddc->coarse_psi_D);CHKERRQ(ierr); 1733f4ddd8eeSStefano Zampini } 1734f4ddd8eeSStefano Zampini } 173506656605SStefano Zampini 1736f4ddd8eeSStefano Zampini if (!pcbddc->coarse_phi_B) { 173706656605SStefano Zampini PetscScalar *marray; 173888ebb749SStefano Zampini 173906656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17408eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 174106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 174288ebb749SStefano Zampini } 17433301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 174406656605SStefano Zampini n *= 2; 174588ebb749SStefano Zampini } 174606656605SStefano Zampini ierr = PetscCalloc1(n,&marray);CHKERRQ(ierr); 174706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray,&pcbddc->coarse_phi_B);CHKERRQ(ierr); 174806656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 17498eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 175006656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_phi_D);CHKERRQ(ierr); 175106656605SStefano Zampini n += n_D*pcbddc->local_primal_size; 175288ebb749SStefano Zampini } 17533301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 175406656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_B);CHKERRQ(ierr); 17558eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 175606656605SStefano Zampini n = n_B*pcbddc->local_primal_size; 175706656605SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marray+n,&pcbddc->coarse_psi_D);CHKERRQ(ierr); 175888ebb749SStefano Zampini } 175988ebb749SStefano Zampini } else { 1760c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);CHKERRQ(ierr); 1761c0553b1fSStefano Zampini pcbddc->coarse_psi_B = pcbddc->coarse_phi_B; 17621b968477SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 1763c0553b1fSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);CHKERRQ(ierr); 1764c0553b1fSStefano Zampini pcbddc->coarse_psi_D = pcbddc->coarse_phi_D; 1765c0553b1fSStefano Zampini } 176688ebb749SStefano Zampini } 176706656605SStefano Zampini } 1768019a44ceSStefano Zampini 176906656605SStefano Zampini /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */ 17704f1b2e48SStefano Zampini p0_lidx_I = NULL; 17714f1b2e48SStefano Zampini if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) { 1772d12edf2fSStefano Zampini const PetscInt *idxs; 1773d12edf2fSStefano Zampini 1774d12edf2fSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 17754f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);CHKERRQ(ierr); 17764f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 17774f1b2e48SStefano Zampini ierr = PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);CHKERRQ(ierr); 17784f1b2e48SStefano Zampini } 1779d12edf2fSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,&idxs);CHKERRQ(ierr); 1780d12edf2fSStefano Zampini } 1781d16cbb6bSStefano Zampini 178206656605SStefano Zampini /* vertices */ 178306656605SStefano Zampini if (n_vertices) { 178416f15bc4SStefano Zampini 1785ffd830a3SStefano Zampini ierr = MatConvert(A_VV,MATDENSE,MAT_REUSE_MATRIX,&A_VV);CHKERRQ(ierr); 178604708bb6SStefano Zampini 178716f15bc4SStefano Zampini if (n_R) { 178814393ed6SStefano Zampini Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */ 178906656605SStefano Zampini PetscBLASInt B_N,B_one = 1; 179016f15bc4SStefano Zampini PetscScalar *x,*y; 179104708bb6SStefano Zampini PetscBool isseqaij; 179206656605SStefano Zampini 179321eccb56SStefano Zampini ierr = MatScale(A_RV,m_one);CHKERRQ(ierr); 179414393ed6SStefano Zampini if (need_benign_correction) { 179514393ed6SStefano Zampini ISLocalToGlobalMapping RtoN; 179614393ed6SStefano Zampini IS is_p0; 179714393ed6SStefano Zampini PetscInt *idxs_p0,n; 179814393ed6SStefano Zampini 179914393ed6SStefano Zampini ierr = PetscMalloc1(pcbddc->benign_n,&idxs_p0);CHKERRQ(ierr); 180014393ed6SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);CHKERRQ(ierr); 180114393ed6SStefano Zampini ierr = ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);CHKERRQ(ierr); 180214393ed6SStefano Zampini if (n != pcbddc->benign_n) { 180314393ed6SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %d != %d\n",n,pcbddc->benign_n); 180414393ed6SStefano Zampini } 180514393ed6SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&RtoN);CHKERRQ(ierr); 180614393ed6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);CHKERRQ(ierr); 180714393ed6SStefano Zampini ierr = MatGetSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);CHKERRQ(ierr); 180814393ed6SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 180914393ed6SStefano Zampini } 181014393ed6SStefano Zampini 1811ffd830a3SStefano Zampini if (lda_rhs == n_R) { 1812ffd830a3SStefano Zampini ierr = MatConvert(A_RV,MATDENSE,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1813ffd830a3SStefano Zampini } else { 1814ca92afb2SStefano Zampini PetscScalar *av,*array; 1815ca92afb2SStefano Zampini const PetscInt *xadj,*adjncy; 1816ca92afb2SStefano Zampini PetscInt n; 1817ca92afb2SStefano Zampini PetscBool flg_row; 1818ffd830a3SStefano Zampini 1819ca92afb2SStefano Zampini array = work+lda_rhs*n_vertices; 1820ca92afb2SStefano Zampini ierr = PetscMemzero(array,lda_rhs*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 1821ca92afb2SStefano Zampini ierr = MatConvert(A_RV,MATSEQAIJ,MAT_REUSE_MATRIX,&A_RV);CHKERRQ(ierr); 1822ca92afb2SStefano Zampini ierr = MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1823ca92afb2SStefano Zampini ierr = MatSeqAIJGetArray(A_RV,&av);CHKERRQ(ierr); 1824ca92afb2SStefano Zampini for (i=0;i<n;i++) { 1825ca92afb2SStefano Zampini PetscInt j; 1826ca92afb2SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j]; 1827ffd830a3SStefano Zampini } 1828ca92afb2SStefano Zampini ierr = MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 1829ca92afb2SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1830ca92afb2SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);CHKERRQ(ierr); 1831ffd830a3SStefano Zampini } 1832ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1833a3df083aSStefano Zampini if (need_benign_correction) { 1834df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1835a3df083aSStefano Zampini PetscScalar *marr; 1836a3df083aSStefano Zampini 1837a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 183814393ed6SStefano Zampini /* need \Phi^T A_RV = (I+L)A_RV, L given by 183914393ed6SStefano Zampini 184014393ed6SStefano Zampini | 0 0 0 | (V) 184114393ed6SStefano Zampini L = | 0 0 -1 | (P-p0) 184214393ed6SStefano Zampini | 0 0 -1 | (p0) 184314393ed6SStefano Zampini 184414393ed6SStefano Zampini */ 1845df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 184614393ed6SStefano Zampini const PetscScalar *vals; 184714393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 184814393ed6SStefano Zampini PetscInt n,j,nz; 184914393ed6SStefano Zampini 1850df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1851df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 185214393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 185314393ed6SStefano Zampini for (j=0;j<n;j++) { 185414393ed6SStefano Zampini PetscScalar val = vals[j]; 185514393ed6SStefano Zampini PetscInt k,col = idxs[j]; 185614393ed6SStefano Zampini for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val; 185714393ed6SStefano Zampini } 185814393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1859df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 186014393ed6SStefano Zampini } 186172b8c272SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 186272b8c272SStefano Zampini } 186372b8c272SStefano Zampini if (F) { 186414393ed6SStefano Zampini /* need to correct the rhs */ 186572b8c272SStefano Zampini if (need_benign_correction) { 186672b8c272SStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 186772b8c272SStefano Zampini PetscScalar *marr; 186872b8c272SStefano Zampini 186972b8c272SStefano Zampini ierr = MatDenseGetArray(A_RV,&marr);CHKERRQ(ierr); 1870a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1871a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1872df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE);CHKERRQ(ierr); 1873a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1874a3df083aSStefano Zampini } 1875a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RV,&marr);CHKERRQ(ierr); 1876a3df083aSStefano Zampini } 187706656605SStefano Zampini ierr = MatMatSolve(F,A_RV,A_RRmA_RV);CHKERRQ(ierr); 187814393ed6SStefano Zampini /* need to correct the solution */ 1879a3df083aSStefano Zampini if (need_benign_correction) { 1880df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 1881a3df083aSStefano Zampini PetscScalar *marr; 1882a3df083aSStefano Zampini 1883a3df083aSStefano Zampini ierr = MatDenseGetArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1884a3df083aSStefano Zampini for (i=0;i<n_vertices;i++) { 1885a3df083aSStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);CHKERRQ(ierr); 1886df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE);CHKERRQ(ierr); 1887a3df083aSStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 1888a3df083aSStefano Zampini } 1889a3df083aSStefano Zampini ierr = MatDenseRestoreArray(A_RRmA_RV,&marr);CHKERRQ(ierr); 1890a3df083aSStefano Zampini } 189106656605SStefano Zampini } else { 189206656605SStefano Zampini ierr = MatDenseGetArray(A_RV,&y);CHKERRQ(ierr); 189306656605SStefano Zampini for (i=0;i<n_vertices;i++) { 1894ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);CHKERRQ(ierr); 1895ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);CHKERRQ(ierr); 189606656605SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 189706656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 189806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 189906656605SStefano Zampini } 190006656605SStefano Zampini ierr = MatDenseRestoreArray(A_RV,&y);CHKERRQ(ierr); 190106656605SStefano Zampini } 190280677318SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 1903ffd830a3SStefano Zampini /* S_VV and S_CV */ 190406656605SStefano Zampini if (n_constraints) { 190506656605SStefano Zampini Mat B; 190680677318SStefano Zampini 1907ffd830a3SStefano Zampini ierr = PetscMemzero(work+lda_rhs*n_vertices,n_B*n_vertices*sizeof(PetscScalar));CHKERRQ(ierr); 190880677318SStefano Zampini for (i=0;i<n_vertices;i++) { 1909ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);CHKERRQ(ierr); 1910ffd830a3SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);CHKERRQ(ierr); 191180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 191280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 191380677318SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 191480677318SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 191580677318SStefano Zampini } 1916ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 191780677318SStefano Zampini ierr = MatMatMult(pcbddc->local_auxmat1,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&S_CV);CHKERRQ(ierr); 191880677318SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 1919ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);CHKERRQ(ierr); 192080677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CV,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 192106656605SStefano Zampini ierr = MatScale(S_CV,m_one);CHKERRQ(ierr); 1922ffd830a3SStefano Zampini ierr = PetscBLASIntCast(lda_rhs*n_vertices,&B_N);CHKERRQ(ierr); 1923ffd830a3SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one)); 192406656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 192506656605SStefano Zampini } 192604708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 192704708bb6SStefano Zampini if (!isseqaij) { /* MatMatMult with SEQ(S)BAIJ below will raise an error */ 192804708bb6SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 192904708bb6SStefano Zampini } 1930ffd830a3SStefano Zampini if (lda_rhs != n_R) { 1931ffd830a3SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 1932ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);CHKERRQ(ierr); 1933ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);CHKERRQ(ierr); 1934ffd830a3SStefano Zampini } 193506656605SStefano Zampini ierr = MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);CHKERRQ(ierr); 193614393ed6SStefano Zampini /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */ 193714393ed6SStefano Zampini if (need_benign_correction) { 1938df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 193914393ed6SStefano Zampini PetscScalar *marr,*sums; 194014393ed6SStefano Zampini 194114393ed6SStefano Zampini ierr = PetscMalloc1(n_vertices,&sums);CHKERRQ(ierr); 194214393ed6SStefano Zampini ierr = MatDenseGetArray(S_VVt,&marr); 1943df4d28bfSStefano Zampini for (i=0;i<reuse_solver->benign_n;i++) { 194414393ed6SStefano Zampini const PetscScalar *vals; 194514393ed6SStefano Zampini const PetscInt *idxs,*idxs_zero; 194614393ed6SStefano Zampini PetscInt n,j,nz; 194714393ed6SStefano Zampini 1948df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);CHKERRQ(ierr); 1949df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 195014393ed6SStefano Zampini for (j=0;j<n_vertices;j++) { 195114393ed6SStefano Zampini PetscInt k; 195214393ed6SStefano Zampini sums[j] = 0.; 195314393ed6SStefano Zampini for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs]; 195414393ed6SStefano Zampini } 195514393ed6SStefano Zampini ierr = MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 195614393ed6SStefano Zampini for (j=0;j<n;j++) { 195714393ed6SStefano Zampini PetscScalar val = vals[j]; 195814393ed6SStefano Zampini PetscInt k; 195914393ed6SStefano Zampini for (k=0;k<n_vertices;k++) { 196014393ed6SStefano Zampini marr[idxs[j]+k*n_vertices] += val*sums[k]; 196114393ed6SStefano Zampini } 196214393ed6SStefano Zampini } 196314393ed6SStefano Zampini ierr = MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);CHKERRQ(ierr); 1964df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);CHKERRQ(ierr); 196514393ed6SStefano Zampini } 196614393ed6SStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 196714393ed6SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&marr); 196814393ed6SStefano Zampini ierr = MatDestroy(&A_RV_bcorr);CHKERRQ(ierr); 196914393ed6SStefano Zampini } 197080677318SStefano Zampini ierr = MatDestroy(&A_RRmA_RV);CHKERRQ(ierr); 197106656605SStefano Zampini ierr = PetscBLASIntCast(n_vertices*n_vertices,&B_N);CHKERRQ(ierr); 197206656605SStefano Zampini ierr = MatDenseGetArray(A_VV,&x);CHKERRQ(ierr); 197306656605SStefano Zampini ierr = MatDenseGetArray(S_VVt,&y);CHKERRQ(ierr); 197406656605SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one)); 197506656605SStefano Zampini ierr = MatDenseRestoreArray(A_VV,&x);CHKERRQ(ierr); 197606656605SStefano Zampini ierr = MatDenseRestoreArray(S_VVt,&y);CHKERRQ(ierr); 197706656605SStefano Zampini ierr = MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1978d16cbb6bSStefano Zampini ierr = MatDestroy(&S_VVt);CHKERRQ(ierr); 1979019a44ceSStefano Zampini } else { 1980d16cbb6bSStefano Zampini ierr = MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 1981d16cbb6bSStefano Zampini } 198221eccb56SStefano Zampini ierr = MatDestroy(&A_VV);CHKERRQ(ierr); 1983d16cbb6bSStefano Zampini 198406656605SStefano Zampini /* coarse basis functions */ 198506656605SStefano Zampini for (i=0;i<n_vertices;i++) { 198616f15bc4SStefano Zampini PetscScalar *y; 198716f15bc4SStefano Zampini 1988ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 198906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 199006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 199106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 199306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 199406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 199506656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 199606656605SStefano Zampini 199706656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 19984f1b2e48SStefano Zampini PetscInt j; 19994f1b2e48SStefano Zampini 200006656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 200106656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 200206656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200306656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 200406656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20054f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 200606656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 200706656605SStefano Zampini } 200806656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 200906656605SStefano Zampini } 201004708bb6SStefano Zampini /* if n_R == 0 the object is not destroyed */ 201104708bb6SStefano Zampini ierr = MatDestroy(&A_RV);CHKERRQ(ierr); 201206656605SStefano Zampini } 201306656605SStefano Zampini 201406656605SStefano Zampini if (n_constraints) { 201506656605SStefano Zampini Mat B; 201606656605SStefano Zampini 2017ffd830a3SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);CHKERRQ(ierr); 201806656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 201980677318SStefano Zampini ierr = MatMatMult(local_auxmat2_R,S_CC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B);CHKERRQ(ierr); 202006656605SStefano Zampini ierr = MatScale(S_CC,m_one);CHKERRQ(ierr); 202106656605SStefano Zampini if (n_vertices) { 202280677318SStefano Zampini if (isCHOL) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */ 202380677318SStefano Zampini ierr = MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);CHKERRQ(ierr); 202480677318SStefano Zampini } else { 202580677318SStefano Zampini Mat S_VCt; 202680677318SStefano Zampini 2027ffd830a3SStefano Zampini if (lda_rhs != n_R) { 2028ffd830a3SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 202972b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);CHKERRQ(ierr); 2030ffd830a3SStefano Zampini ierr = MatSeqDenseSetLDA(B,lda_rhs);CHKERRQ(ierr); 2031ffd830a3SStefano Zampini } 203280677318SStefano Zampini ierr = MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);CHKERRQ(ierr); 203380677318SStefano Zampini ierr = MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 203480677318SStefano Zampini ierr = MatDestroy(&S_VCt);CHKERRQ(ierr); 203580677318SStefano Zampini } 203606656605SStefano Zampini } 203706656605SStefano Zampini ierr = MatDestroy(&B);CHKERRQ(ierr); 203806656605SStefano Zampini /* coarse basis functions */ 203906656605SStefano Zampini for (i=0;i<n_constraints;i++) { 204006656605SStefano Zampini PetscScalar *y; 204106656605SStefano Zampini 2042ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);CHKERRQ(ierr); 204306656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 204406656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));CHKERRQ(ierr); 204506656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 204606656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 204706656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);CHKERRQ(ierr); 204806656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 204906656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 20504f1b2e48SStefano Zampini PetscInt j; 20514f1b2e48SStefano Zampini 205206656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 205306656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));CHKERRQ(ierr); 205406656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205506656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 205606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 20574f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0; 205806656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);CHKERRQ(ierr); 205906656605SStefano Zampini } 206006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 206106656605SStefano Zampini } 206206656605SStefano Zampini } 206380677318SStefano Zampini if (n_constraints) { 206480677318SStefano Zampini ierr = MatDestroy(&local_auxmat2_R);CHKERRQ(ierr); 206580677318SStefano Zampini } 20664f1b2e48SStefano Zampini ierr = PetscFree(p0_lidx_I);CHKERRQ(ierr); 206772b8c272SStefano Zampini 206872b8c272SStefano Zampini /* coarse matrix entries relative to B_0 */ 206972b8c272SStefano Zampini if (pcbddc->benign_n) { 207072b8c272SStefano Zampini Mat B0_B,B0_BPHI; 207172b8c272SStefano Zampini IS is_dummy; 207272b8c272SStefano Zampini PetscScalar *data; 207372b8c272SStefano Zampini PetscInt j; 207472b8c272SStefano Zampini 207572b8c272SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 207672b8c272SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 207772b8c272SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 207872b8c272SStefano Zampini ierr = MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 207972b8c272SStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 208072b8c272SStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data);CHKERRQ(ierr); 208172b8c272SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 208272b8c272SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 208372b8c272SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 208472b8c272SStefano Zampini coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j]; 208572b8c272SStefano Zampini coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j]; 208672b8c272SStefano Zampini } 208772b8c272SStefano Zampini } 208872b8c272SStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data);CHKERRQ(ierr); 208972b8c272SStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 209072b8c272SStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 209172b8c272SStefano Zampini } 2092019a44ceSStefano Zampini 209306656605SStefano Zampini /* compute other basis functions for non-symmetric problems */ 20943301b35fSStefano Zampini if (!pcbddc->symmetric_primal) { 2095ffd830a3SStefano Zampini Mat B_V=NULL,B_C=NULL; 2096ffd830a3SStefano Zampini PetscScalar *marray; 209706656605SStefano Zampini 209806656605SStefano Zampini if (n_constraints) { 2099ffd830a3SStefano Zampini Mat S_CCT,C_CRT; 210006656605SStefano Zampini 2101ffd830a3SStefano Zampini ierr = MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);CHKERRQ(ierr); 210206656605SStefano Zampini ierr = MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);CHKERRQ(ierr); 2103ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);CHKERRQ(ierr); 210416f15bc4SStefano Zampini ierr = MatDestroy(&S_CCT);CHKERRQ(ierr); 210506656605SStefano Zampini if (n_vertices) { 2106ffd830a3SStefano Zampini Mat S_VCT; 210706656605SStefano Zampini 210806656605SStefano Zampini ierr = MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);CHKERRQ(ierr); 2109ffd830a3SStefano Zampini ierr = MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);CHKERRQ(ierr); 211016f15bc4SStefano Zampini ierr = MatDestroy(&S_VCT);CHKERRQ(ierr); 211106656605SStefano Zampini } 2112ffd830a3SStefano Zampini ierr = MatDestroy(&C_CRT);CHKERRQ(ierr); 211306656605SStefano Zampini } 211416f15bc4SStefano Zampini if (n_vertices && n_R) { 2115ffd830a3SStefano Zampini PetscScalar *av,*marray; 2116ffd830a3SStefano Zampini const PetscInt *xadj,*adjncy; 2117ffd830a3SStefano Zampini PetscInt n; 2118ffd830a3SStefano Zampini PetscBool flg_row; 211906656605SStefano Zampini 2120ffd830a3SStefano Zampini /* B_V = B_V - A_VR^T */ 2121ffd830a3SStefano Zampini ierr = MatConvert(A_VR,MATSEQAIJ,MAT_REUSE_MATRIX,&A_VR);CHKERRQ(ierr); 2122ffd830a3SStefano Zampini ierr = MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2123ffd830a3SStefano Zampini ierr = MatSeqAIJGetArray(A_VR,&av);CHKERRQ(ierr); 2124ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2125ffd830a3SStefano Zampini for (i=0;i<n;i++) { 2126ffd830a3SStefano Zampini PetscInt j; 2127ffd830a3SStefano Zampini for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j]; 2128ffd830a3SStefano Zampini } 2129ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2130ffd830a3SStefano Zampini ierr = MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 2131ffd830a3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 213206656605SStefano Zampini } 213306656605SStefano Zampini 2134ffd830a3SStefano Zampini /* currently there's no support for MatTransposeMatSolve(F,B,X) */ 2135ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_V,&marray);CHKERRQ(ierr); 2136ffd830a3SStefano Zampini for (i=0;i<n_vertices;i++) { 2137ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);CHKERRQ(ierr); 2138ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 213906656605SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 214006656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 214106656605SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 214206656605SStefano Zampini } 2143ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_V,&marray);CHKERRQ(ierr); 2144ffd830a3SStefano Zampini ierr = MatDenseGetArray(B_C,&marray);CHKERRQ(ierr); 2145ffd830a3SStefano Zampini for (i=n_vertices;i<n_constraints+n_vertices;i++) { 2146ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);CHKERRQ(ierr); 2147ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec2_R,work+i*n_R);CHKERRQ(ierr); 2148ffd830a3SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 2149ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 2150ffd830a3SStefano Zampini ierr = VecResetArray(pcbddc->vec2_R);CHKERRQ(ierr); 215106656605SStefano Zampini } 2152ffd830a3SStefano Zampini ierr = MatDenseRestoreArray(B_C,&marray);CHKERRQ(ierr); 215306656605SStefano Zampini /* coarse basis functions */ 215406656605SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 215506656605SStefano Zampini PetscScalar *y; 215606656605SStefano Zampini 2157ffd830a3SStefano Zampini ierr = VecPlaceArray(pcbddc->vec1_R,work+i*n_R);CHKERRQ(ierr); 215806656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 215906656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_B,y+n_B*i);CHKERRQ(ierr); 216006656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 216106656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 216206656605SStefano Zampini if (i<n_vertices) { 216306656605SStefano Zampini y[n_B*i+idx_V_B[i]] = 1.0; 216406656605SStefano Zampini } 216506656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);CHKERRQ(ierr); 216606656605SStefano Zampini ierr = VecResetArray(pcis->vec1_B);CHKERRQ(ierr); 216706656605SStefano Zampini 216806656605SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 216906656605SStefano Zampini ierr = MatDenseGetArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 217006656605SStefano Zampini ierr = VecPlaceArray(pcis->vec1_D,y+n_D*i);CHKERRQ(ierr); 217106656605SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217206656605SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 217306656605SStefano Zampini ierr = VecResetArray(pcis->vec1_D);CHKERRQ(ierr); 217406656605SStefano Zampini ierr = MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);CHKERRQ(ierr); 217506656605SStefano Zampini } 217606656605SStefano Zampini ierr = VecResetArray(pcbddc->vec1_R);CHKERRQ(ierr); 217706656605SStefano Zampini } 2178ffd830a3SStefano Zampini ierr = MatDestroy(&B_V);CHKERRQ(ierr); 2179ffd830a3SStefano Zampini ierr = MatDestroy(&B_C);CHKERRQ(ierr); 218006656605SStefano Zampini } 2181d62866d3SStefano Zampini /* free memory */ 218288ebb749SStefano Zampini ierr = PetscFree(idx_V_B);CHKERRQ(ierr); 218306656605SStefano Zampini ierr = MatDestroy(&S_VV);CHKERRQ(ierr); 218406656605SStefano Zampini ierr = MatDestroy(&S_CV);CHKERRQ(ierr); 218506656605SStefano Zampini ierr = MatDestroy(&S_VC);CHKERRQ(ierr); 218606656605SStefano Zampini ierr = MatDestroy(&S_CC);CHKERRQ(ierr); 2187d62866d3SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 2188d62866d3SStefano Zampini if (n_vertices) { 2189d62866d3SStefano Zampini ierr = MatDestroy(&A_VR);CHKERRQ(ierr); 2190d62866d3SStefano Zampini } 2191d62866d3SStefano Zampini if (n_constraints) { 2192d62866d3SStefano Zampini ierr = MatDestroy(&C_CR);CHKERRQ(ierr); 2193d62866d3SStefano Zampini } 219488ebb749SStefano Zampini /* Checking coarse_sub_mat and coarse basis functios */ 219588ebb749SStefano Zampini /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 219688ebb749SStefano Zampini /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */ 2197d12edf2fSStefano Zampini if (pcbddc->dbg_flag) { 219888ebb749SStefano Zampini Mat coarse_sub_mat; 219925084f0cSStefano Zampini Mat AUXMAT,TM1,TM2,TM3,TM4; 220088ebb749SStefano Zampini Mat coarse_phi_D,coarse_phi_B; 220188ebb749SStefano Zampini Mat coarse_psi_D,coarse_psi_B; 220288ebb749SStefano Zampini Mat A_II,A_BB,A_IB,A_BI; 22038bec7fa6SStefano Zampini Mat C_B,CPHI; 22048bec7fa6SStefano Zampini IS is_dummy; 22058bec7fa6SStefano Zampini Vec mones; 220688ebb749SStefano Zampini MatType checkmattype=MATSEQAIJ; 220788ebb749SStefano Zampini PetscReal real_value; 220888ebb749SStefano Zampini 2209a3df083aSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2210a3df083aSStefano Zampini Mat A; 2211a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,NULL,NULL,&A);CHKERRQ(ierr); 2212a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 2213a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 2214a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 2215a3df083aSStefano Zampini ierr = MatGetSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2216a3df083aSStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 2217a3df083aSStefano Zampini } else { 221888ebb749SStefano Zampini ierr = MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);CHKERRQ(ierr); 221988ebb749SStefano Zampini ierr = MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 222088ebb749SStefano Zampini ierr = MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 222188ebb749SStefano Zampini ierr = MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 2222a3df083aSStefano Zampini } 222388ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);CHKERRQ(ierr); 222488ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);CHKERRQ(ierr); 2225ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 222688ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);CHKERRQ(ierr); 222788ebb749SStefano Zampini ierr = MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);CHKERRQ(ierr); 222888ebb749SStefano Zampini } 222988ebb749SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);CHKERRQ(ierr); 223088ebb749SStefano Zampini 223125084f0cSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 22323301b35fSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);CHKERRQ(ierr); 223325084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2234ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 223588ebb749SStefano Zampini ierr = MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 223688ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 223788ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 223888ebb749SStefano Zampini ierr = MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 223988ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 224088ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 224188ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 224288ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 224388ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 224488ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 224588ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 224688ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 224788ebb749SStefano Zampini } else { 224888ebb749SStefano Zampini ierr = MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);CHKERRQ(ierr); 224988ebb749SStefano Zampini ierr = MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);CHKERRQ(ierr); 225088ebb749SStefano Zampini ierr = MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 225188ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);CHKERRQ(ierr); 225288ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 225388ebb749SStefano Zampini ierr = MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);CHKERRQ(ierr); 225488ebb749SStefano Zampini ierr = MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);CHKERRQ(ierr); 225588ebb749SStefano Zampini ierr = MatDestroy(&AUXMAT);CHKERRQ(ierr); 225688ebb749SStefano Zampini } 225788ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 225888ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 225988ebb749SStefano Zampini ierr = MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 226088ebb749SStefano Zampini ierr = MatConvert(TM1,MATSEQDENSE,MAT_REUSE_MATRIX,&TM1);CHKERRQ(ierr); 22614f1b2e48SStefano Zampini if (pcbddc->benign_n) { 2262fc227af8SStefano Zampini Mat B0_B,B0_BPHI; 2263d12edf2fSStefano Zampini PetscScalar *data,*data2; 22644f1b2e48SStefano Zampini PetscInt j; 2265d12edf2fSStefano Zampini 22664f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 2267fc227af8SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 2268d12edf2fSStefano Zampini ierr = MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);CHKERRQ(ierr); 2269d12edf2fSStefano Zampini ierr = MatConvert(B0_BPHI,MATSEQDENSE,MAT_REUSE_MATRIX,&B0_BPHI);CHKERRQ(ierr); 2270d12edf2fSStefano Zampini ierr = MatDenseGetArray(TM1,&data);CHKERRQ(ierr); 2271d12edf2fSStefano Zampini ierr = MatDenseGetArray(B0_BPHI,&data2);CHKERRQ(ierr); 22724f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) { 22734f1b2e48SStefano Zampini PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j; 2274d12edf2fSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 22754f1b2e48SStefano Zampini data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j]; 22764f1b2e48SStefano Zampini data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j]; 22774f1b2e48SStefano Zampini } 2278d12edf2fSStefano Zampini } 2279d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(TM1,&data);CHKERRQ(ierr); 2280d12edf2fSStefano Zampini ierr = MatDenseRestoreArray(B0_BPHI,&data2);CHKERRQ(ierr); 2281d12edf2fSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 2282d12edf2fSStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 2283d12edf2fSStefano Zampini ierr = MatDestroy(&B0_BPHI);CHKERRQ(ierr); 2284d12edf2fSStefano Zampini } 2285d12edf2fSStefano Zampini #if 0 2286d12edf2fSStefano Zampini { 2287d12edf2fSStefano Zampini PetscViewer viewer; 2288d12edf2fSStefano Zampini char filename[256]; 2289ffd830a3SStefano Zampini sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level); 2290d12edf2fSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);CHKERRQ(ierr); 2291d12edf2fSStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2292ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");CHKERRQ(ierr); 2293ffd830a3SStefano Zampini ierr = MatView(coarse_sub_mat,viewer);CHKERRQ(ierr); 2294ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)TM1,"projected");CHKERRQ(ierr); 2295d12edf2fSStefano Zampini ierr = MatView(TM1,viewer);CHKERRQ(ierr); 229672b8c272SStefano Zampini if (save_change) { 229772b8c272SStefano Zampini Mat phi_B; 229872b8c272SStefano Zampini ierr = MatMatMult(save_change,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&phi_B);CHKERRQ(ierr); 229972b8c272SStefano Zampini ierr = PetscObjectSetName((PetscObject)phi_B,"phi_B");CHKERRQ(ierr); 230072b8c272SStefano Zampini ierr = MatView(phi_B,viewer);CHKERRQ(ierr); 230172b8c272SStefano Zampini ierr = MatDestroy(&phi_B);CHKERRQ(ierr); 230272b8c272SStefano Zampini } else { 2303ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");CHKERRQ(ierr); 2304ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_B,viewer);CHKERRQ(ierr); 230572b8c272SStefano Zampini } 2306ffd830a3SStefano Zampini if (pcbddc->coarse_phi_D) { 2307ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");CHKERRQ(ierr); 2308ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_phi_D,viewer);CHKERRQ(ierr); 2309ffd830a3SStefano Zampini } 2310ffd830a3SStefano Zampini if (pcbddc->coarse_psi_B) { 2311ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");CHKERRQ(ierr); 2312ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_B,viewer);CHKERRQ(ierr); 2313ffd830a3SStefano Zampini } 231472b8c272SStefano Zampini if (pcbddc->coarse_psi_D) { 2315ffd830a3SStefano Zampini ierr = PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");CHKERRQ(ierr); 2316ffd830a3SStefano Zampini ierr = MatView(pcbddc->coarse_psi_D,viewer);CHKERRQ(ierr); 2317ffd830a3SStefano Zampini } 2318d12edf2fSStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 2319d12edf2fSStefano Zampini } 2320d12edf2fSStefano Zampini #endif 232181d9aea3SBarry Smith ierr = MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr); 23228bec7fa6SStefano Zampini ierr = MatNorm(TM1,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 23231575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 232406656605SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 23258bec7fa6SStefano Zampini 23268bec7fa6SStefano Zampini /* check constraints */ 23274f1b2e48SStefano Zampini if (!pcbddc->benign_n) { /* TODO: add benign case */ 23288bec7fa6SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size,0,1,&is_dummy);CHKERRQ(ierr); 23298bec7fa6SStefano Zampini ierr = MatGetSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B); 23308bec7fa6SStefano Zampini ierr = MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 23318bec7fa6SStefano Zampini ierr = MatCreateVecs(CPHI,&mones,NULL);CHKERRQ(ierr); 23328bec7fa6SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 23338bec7fa6SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 23348bec7fa6SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2335bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 2336ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 2337bdae7319SStefano Zampini ierr = MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);CHKERRQ(ierr); 2338bdae7319SStefano Zampini ierr = VecSet(mones,-1.0);CHKERRQ(ierr); 2339bdae7319SStefano Zampini ierr = MatDiagonalSet(CPHI,mones,ADD_VALUES);CHKERRQ(ierr); 2340bdae7319SStefano Zampini ierr = MatNorm(CPHI,NORM_FROBENIUS,&real_value);CHKERRQ(ierr); 2341bdae7319SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);CHKERRQ(ierr); 234288ebb749SStefano Zampini } 23438bec7fa6SStefano Zampini ierr = MatDestroy(&C_B);CHKERRQ(ierr); 23448bec7fa6SStefano Zampini ierr = MatDestroy(&CPHI);CHKERRQ(ierr); 23458bec7fa6SStefano Zampini ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 23468bec7fa6SStefano Zampini ierr = VecDestroy(&mones);CHKERRQ(ierr); 2347d12edf2fSStefano Zampini } 234825084f0cSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 234988ebb749SStefano Zampini ierr = MatDestroy(&A_II);CHKERRQ(ierr); 235088ebb749SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 235188ebb749SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 235288ebb749SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 235388ebb749SStefano Zampini ierr = MatDestroy(&TM1);CHKERRQ(ierr); 235488ebb749SStefano Zampini ierr = MatDestroy(&TM2);CHKERRQ(ierr); 235588ebb749SStefano Zampini ierr = MatDestroy(&TM3);CHKERRQ(ierr); 235688ebb749SStefano Zampini ierr = MatDestroy(&TM4);CHKERRQ(ierr); 235788ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_D);CHKERRQ(ierr); 235888ebb749SStefano Zampini ierr = MatDestroy(&coarse_phi_B);CHKERRQ(ierr); 2359ffd830a3SStefano Zampini if (!pcbddc->symmetric_primal) { 236088ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_D);CHKERRQ(ierr); 236188ebb749SStefano Zampini ierr = MatDestroy(&coarse_psi_B);CHKERRQ(ierr); 236288ebb749SStefano Zampini } 236388ebb749SStefano Zampini ierr = MatDestroy(&coarse_sub_mat);CHKERRQ(ierr); 236488ebb749SStefano Zampini } 23658629588bSStefano Zampini /* get back data */ 23668629588bSStefano Zampini *coarse_submat_vals_n = coarse_submat_vals; 236788ebb749SStefano Zampini PetscFunctionReturn(0); 236888ebb749SStefano Zampini } 236988ebb749SStefano Zampini 237088ebb749SStefano Zampini #undef __FUNCT__ 2371d65f70fdSStefano Zampini #define __FUNCT__ "MatGetSubMatrixUnsorted" 2372d648f858SStefano Zampini PetscErrorCode MatGetSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B) 2373aa0d41d4SStefano Zampini { 2374d65f70fdSStefano Zampini Mat *work_mat; 2375d65f70fdSStefano Zampini IS isrow_s,iscol_s; 2376d65f70fdSStefano Zampini PetscBool rsorted,csorted; 2377d65f70fdSStefano Zampini PetscInt rsize,*idxs_perm_r,csize,*idxs_perm_c; 2378aa0d41d4SStefano Zampini PetscErrorCode ierr; 2379aa0d41d4SStefano Zampini 2380aa0d41d4SStefano Zampini PetscFunctionBegin; 2381d65f70fdSStefano Zampini ierr = ISSorted(isrow,&rsorted);CHKERRQ(ierr); 2382d65f70fdSStefano Zampini ierr = ISSorted(iscol,&csorted);CHKERRQ(ierr); 2383d65f70fdSStefano Zampini ierr = ISGetLocalSize(isrow,&rsize);CHKERRQ(ierr); 2384d65f70fdSStefano Zampini ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2385aa0d41d4SStefano Zampini 2386d65f70fdSStefano Zampini if (!rsorted) { 2387906d46d4SStefano Zampini const PetscInt *idxs; 2388906d46d4SStefano Zampini PetscInt *idxs_sorted,i; 2389aa0d41d4SStefano Zampini 2390d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_perm_r);CHKERRQ(ierr); 2391d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_sorted);CHKERRQ(ierr); 2392d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2393d65f70fdSStefano Zampini idxs_perm_r[i] = i; 2394aa0d41d4SStefano Zampini } 2395d65f70fdSStefano Zampini ierr = ISGetIndices(isrow,&idxs);CHKERRQ(ierr); 2396d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);CHKERRQ(ierr); 2397d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2398d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_r[i]]; 2399aa0d41d4SStefano Zampini } 2400d65f70fdSStefano Zampini ierr = ISRestoreIndices(isrow,&idxs);CHKERRQ(ierr); 2401d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);CHKERRQ(ierr); 2402d65f70fdSStefano Zampini } else { 2403d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 2404d65f70fdSStefano Zampini isrow_s = isrow; 2405aa0d41d4SStefano Zampini } 2406906d46d4SStefano Zampini 2407d65f70fdSStefano Zampini if (!csorted) { 2408d65f70fdSStefano Zampini if (isrow == iscol) { 2409d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)isrow_s);CHKERRQ(ierr); 2410d65f70fdSStefano Zampini iscol_s = isrow_s; 2411d65f70fdSStefano Zampini } else { 2412d65f70fdSStefano Zampini const PetscInt *idxs; 2413d65f70fdSStefano Zampini PetscInt *idxs_sorted,i; 2414906d46d4SStefano Zampini 2415d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_perm_c);CHKERRQ(ierr); 2416d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_sorted);CHKERRQ(ierr); 2417d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2418d65f70fdSStefano Zampini idxs_perm_c[i] = i; 2419d65f70fdSStefano Zampini } 2420d65f70fdSStefano Zampini ierr = ISGetIndices(iscol,&idxs);CHKERRQ(ierr); 2421d65f70fdSStefano Zampini ierr = PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);CHKERRQ(ierr); 2422d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2423d65f70fdSStefano Zampini idxs_sorted[i] = idxs[idxs_perm_c[i]]; 2424d65f70fdSStefano Zampini } 2425d65f70fdSStefano Zampini ierr = ISRestoreIndices(iscol,&idxs);CHKERRQ(ierr); 2426d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);CHKERRQ(ierr); 2427d65f70fdSStefano Zampini } 2428d65f70fdSStefano Zampini } else { 2429d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 2430d65f70fdSStefano Zampini iscol_s = iscol; 2431d65f70fdSStefano Zampini } 2432d65f70fdSStefano Zampini 2433d648f858SStefano Zampini ierr = MatGetSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 2434d65f70fdSStefano Zampini 2435d65f70fdSStefano Zampini if (!rsorted || !csorted) { 2436906d46d4SStefano Zampini Mat new_mat; 2437d65f70fdSStefano Zampini IS is_perm_r,is_perm_c; 2438906d46d4SStefano Zampini 2439d65f70fdSStefano Zampini if (!rsorted) { 2440d65f70fdSStefano Zampini PetscInt *idxs_r,i; 2441d65f70fdSStefano Zampini ierr = PetscMalloc1(rsize,&idxs_r);CHKERRQ(ierr); 2442d65f70fdSStefano Zampini for (i=0;i<rsize;i++) { 2443d65f70fdSStefano Zampini idxs_r[idxs_perm_r[i]] = i; 2444906d46d4SStefano Zampini } 2445d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_r);CHKERRQ(ierr); 2446d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);CHKERRQ(ierr); 2447d65f70fdSStefano Zampini } else { 2448d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);CHKERRQ(ierr); 2449906d46d4SStefano Zampini } 2450d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_r);CHKERRQ(ierr); 2451d65f70fdSStefano Zampini 2452d65f70fdSStefano Zampini if (!csorted) { 2453d65f70fdSStefano Zampini if (isrow_s == iscol_s) { 2454d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)is_perm_r);CHKERRQ(ierr); 2455d65f70fdSStefano Zampini is_perm_c = is_perm_r; 2456d65f70fdSStefano Zampini } else { 2457d65f70fdSStefano Zampini PetscInt *idxs_c,i; 2458d65f70fdSStefano Zampini ierr = PetscMalloc1(csize,&idxs_c);CHKERRQ(ierr); 2459d65f70fdSStefano Zampini for (i=0;i<csize;i++) { 2460d65f70fdSStefano Zampini idxs_c[idxs_perm_c[i]] = i; 2461d65f70fdSStefano Zampini } 2462d65f70fdSStefano Zampini ierr = PetscFree(idxs_perm_c);CHKERRQ(ierr); 2463d65f70fdSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);CHKERRQ(ierr); 2464d65f70fdSStefano Zampini } 2465d65f70fdSStefano Zampini } else { 2466d65f70fdSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);CHKERRQ(ierr); 2467d65f70fdSStefano Zampini } 2468d65f70fdSStefano Zampini ierr = ISSetPermutation(is_perm_c);CHKERRQ(ierr); 2469d65f70fdSStefano Zampini 2470d65f70fdSStefano Zampini ierr = MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);CHKERRQ(ierr); 2471d65f70fdSStefano Zampini ierr = MatDestroy(&work_mat[0]);CHKERRQ(ierr); 2472d65f70fdSStefano Zampini work_mat[0] = new_mat; 2473d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_r);CHKERRQ(ierr); 2474d65f70fdSStefano Zampini ierr = ISDestroy(&is_perm_c);CHKERRQ(ierr); 2475d65f70fdSStefano Zampini } 2476d65f70fdSStefano Zampini 2477d65f70fdSStefano Zampini ierr = PetscObjectReference((PetscObject)work_mat[0]);CHKERRQ(ierr); 2478d65f70fdSStefano Zampini *B = work_mat[0]; 2479d65f70fdSStefano Zampini ierr = MatDestroyMatrices(1,&work_mat);CHKERRQ(ierr); 2480d65f70fdSStefano Zampini ierr = ISDestroy(&isrow_s);CHKERRQ(ierr); 2481d65f70fdSStefano Zampini ierr = ISDestroy(&iscol_s);CHKERRQ(ierr); 2482d65f70fdSStefano Zampini PetscFunctionReturn(0); 2483d65f70fdSStefano Zampini } 2484d65f70fdSStefano Zampini 2485d65f70fdSStefano Zampini #undef __FUNCT__ 24865e8657edSStefano Zampini #define __FUNCT__ "PCBDDCComputeLocalMatrix" 24875e8657edSStefano Zampini PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix) 2488aa0d41d4SStefano Zampini { 2489aa0d41d4SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 24905e8657edSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2491d65f70fdSStefano Zampini Mat new_mat; 24925e8657edSStefano Zampini IS is_local,is_global; 2493d65f70fdSStefano Zampini PetscInt local_size; 2494d65f70fdSStefano Zampini PetscBool isseqaij; 2495aa0d41d4SStefano Zampini PetscErrorCode ierr; 2496aa0d41d4SStefano Zampini 2497aa0d41d4SStefano Zampini PetscFunctionBegin; 2498aa0d41d4SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 24995e8657edSStefano Zampini ierr = MatGetSize(matis->A,&local_size,NULL);CHKERRQ(ierr); 25005e8657edSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);CHKERRQ(ierr); 2501b087196eSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);CHKERRQ(ierr); 2502aa0d41d4SStefano Zampini ierr = ISDestroy(&is_local);CHKERRQ(ierr); 2503d648f858SStefano Zampini ierr = MatGetSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);CHKERRQ(ierr); 2504aa0d41d4SStefano Zampini ierr = ISDestroy(&is_global);CHKERRQ(ierr); 2505906d46d4SStefano Zampini 2506906d46d4SStefano Zampini /* check */ 2507906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 2508906d46d4SStefano Zampini Vec x,x_change; 2509906d46d4SStefano Zampini PetscReal error; 2510906d46d4SStefano Zampini 25115e8657edSStefano Zampini ierr = MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);CHKERRQ(ierr); 2512906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 25135e8657edSStefano Zampini ierr = MatMult(ChangeOfBasisMatrix,x,x_change);CHKERRQ(ierr); 2514e176bc59SStefano Zampini ierr = VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2515e176bc59SStefano Zampini ierr = VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2516d65f70fdSStefano Zampini ierr = MatMult(new_mat,matis->x,matis->y);CHKERRQ(ierr); 2517e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2518e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 2519906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 2520906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 2521906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2522906d46d4SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change on N: %1.6e\n",error);CHKERRQ(ierr); 2523906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 2524906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 2525906d46d4SStefano Zampini } 2526906d46d4SStefano Zampini 252722d5777bSStefano Zampini /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */ 25289b28b3ffSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 252922d5777bSStefano Zampini if (isseqaij) { 25301cf9b237SStefano Zampini Mat M; 25311cf9b237SStefano Zampini 25321cf9b237SStefano Zampini ierr = MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 25331cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 25341cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2535aa0d41d4SStefano Zampini } else { 25361cf9b237SStefano Zampini Mat work_mat,M; 25371cf9b237SStefano Zampini 2538aa0d41d4SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);CHKERRQ(ierr); 25391cf9b237SStefano Zampini ierr = MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&M);CHKERRQ(ierr); 25401cf9b237SStefano Zampini ierr = MatSeqAIJCompress(M,&pcbddc->local_mat);CHKERRQ(ierr); 25411cf9b237SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 2542aa0d41d4SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 2543aa0d41d4SStefano Zampini } 25443301b35fSStefano Zampini if (matis->A->symmetric_set) { 25453301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);CHKERRQ(ierr); 2546e496cd5dSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 25473301b35fSStefano Zampini ierr = MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);CHKERRQ(ierr); 2548e496cd5dSStefano Zampini #endif 25493301b35fSStefano Zampini } 255045a1bb75SStefano Zampini /* 255145a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 2552d65f70fdSStefano Zampini ierr = MatView(new_mat,(PetscViewer)0);CHKERRQ(ierr); 255345a1bb75SStefano Zampini */ 2554d65f70fdSStefano Zampini ierr = MatDestroy(&new_mat);CHKERRQ(ierr); 2555aa0d41d4SStefano Zampini PetscFunctionReturn(0); 2556aa0d41d4SStefano Zampini } 2557aa0d41d4SStefano Zampini 2558aa0d41d4SStefano Zampini #undef __FUNCT__ 2559a64d13efSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalScatters" 25608ce42a96SStefano Zampini PetscErrorCode PCBDDCSetUpLocalScatters(PC pc) 2561a64d13efSStefano Zampini { 2562a64d13efSStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 2563a64d13efSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 2564d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 256553892102SStefano Zampini PetscInt *idx_R_local=NULL; 25663a50541eSStefano Zampini PetscInt n_vertices,i,j,n_R,n_D,n_B; 25673a50541eSStefano Zampini PetscInt vbs,bs; 25686816873aSStefano Zampini PetscBT bitmask=NULL; 2569a64d13efSStefano Zampini PetscErrorCode ierr; 2570a64d13efSStefano Zampini 2571a64d13efSStefano Zampini PetscFunctionBegin; 2572b23d619eSStefano Zampini /* 2573b23d619eSStefano Zampini No need to setup local scatters if 2574b23d619eSStefano Zampini - primal space is unchanged 2575b23d619eSStefano Zampini AND 2576b23d619eSStefano Zampini - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains) 2577b23d619eSStefano Zampini AND 2578b23d619eSStefano Zampini - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine 2579b23d619eSStefano Zampini */ 2580b23d619eSStefano Zampini if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) { 2581f4ddd8eeSStefano Zampini PetscFunctionReturn(0); 2582f4ddd8eeSStefano Zampini } 2583f4ddd8eeSStefano Zampini /* destroy old objects */ 2584f4ddd8eeSStefano Zampini ierr = ISDestroy(&pcbddc->is_R_local);CHKERRQ(ierr); 2585f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_B);CHKERRQ(ierr); 2586f4ddd8eeSStefano Zampini ierr = VecScatterDestroy(&pcbddc->R_to_D);CHKERRQ(ierr); 2587a64d13efSStefano Zampini /* Set Non-overlapping dimensions */ 2588b371cd4fSStefano Zampini n_B = pcis->n_B; 2589b371cd4fSStefano Zampini n_D = pcis->n - n_B; 2590b371cd4fSStefano Zampini n_vertices = pcbddc->n_vertices; 25913a50541eSStefano Zampini 2592a64d13efSStefano Zampini /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */ 25936816873aSStefano Zampini 259453892102SStefano Zampini /* create auxiliary bitmask and allocate workspace */ 2595df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2596854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n-n_vertices,&idx_R_local);CHKERRQ(ierr); 2597a64d13efSStefano Zampini ierr = PetscBTCreate(pcis->n,&bitmask);CHKERRQ(ierr); 2598a64d13efSStefano Zampini for (i=0;i<n_vertices;i++) { 25990e6343abSStefano Zampini ierr = PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);CHKERRQ(ierr); 2600a64d13efSStefano Zampini } 2601a64d13efSStefano Zampini 2602a64d13efSStefano Zampini for (i=0, n_R=0; i<pcis->n; i++) { 26034641a718SStefano Zampini if (!PetscBTLookup(bitmask,i)) { 26046816873aSStefano Zampini idx_R_local[n_R++] = i; 2605a64d13efSStefano Zampini } 2606a64d13efSStefano Zampini } 2607df4d28bfSStefano Zampini } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */ 2608df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 26096816873aSStefano Zampini 2610df4d28bfSStefano Zampini ierr = ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2611df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_R,&n_R);CHKERRQ(ierr); 26126816873aSStefano Zampini } 26133a50541eSStefano Zampini 26143a50541eSStefano Zampini /* Block code */ 26153a50541eSStefano Zampini vbs = 1; 26163a50541eSStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&bs);CHKERRQ(ierr); 26173a50541eSStefano Zampini if (bs>1 && !(n_vertices%bs)) { 26183a50541eSStefano Zampini PetscBool is_blocked = PETSC_TRUE; 26193a50541eSStefano Zampini PetscInt *vary; 2620df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 2621785e854fSJed Brown ierr = PetscMalloc1(pcis->n/bs,&vary);CHKERRQ(ierr); 26223a50541eSStefano Zampini ierr = PetscMemzero(vary,pcis->n/bs*sizeof(PetscInt));CHKERRQ(ierr); 2623d3df7717SStefano Zampini /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */ 2624d3df7717SStefano 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 */ 26250e6343abSStefano Zampini for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++; 2626d3df7717SStefano Zampini for (i=0; i<pcis->n/bs; i++) { 26273a50541eSStefano Zampini if (vary[i]!=0 && vary[i]!=bs) { 26283a50541eSStefano Zampini is_blocked = PETSC_FALSE; 26293a50541eSStefano Zampini break; 26303a50541eSStefano Zampini } 26313a50541eSStefano Zampini } 2632d3df7717SStefano Zampini ierr = PetscFree(vary);CHKERRQ(ierr); 2633d3df7717SStefano Zampini } else { 2634d3df7717SStefano Zampini /* Verify directly the R set */ 2635d3df7717SStefano Zampini for (i=0; i<n_R/bs; i++) { 2636d3df7717SStefano Zampini PetscInt j,node=idx_R_local[bs*i]; 2637d3df7717SStefano Zampini for (j=1; j<bs; j++) { 2638d3df7717SStefano Zampini if (node != idx_R_local[bs*i+j]-j) { 2639d3df7717SStefano Zampini is_blocked = PETSC_FALSE; 2640d3df7717SStefano Zampini break; 2641d3df7717SStefano Zampini } 2642d3df7717SStefano Zampini } 2643d3df7717SStefano Zampini } 2644d3df7717SStefano Zampini } 26453a50541eSStefano Zampini if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */ 26463a50541eSStefano Zampini vbs = bs; 26473a50541eSStefano Zampini for (i=0;i<n_R/vbs;i++) { 26483a50541eSStefano Zampini idx_R_local[i] = idx_R_local[vbs*i]/vbs; 26493a50541eSStefano Zampini } 26503a50541eSStefano Zampini } 26513a50541eSStefano Zampini } 26523a50541eSStefano Zampini ierr = ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);CHKERRQ(ierr); 2653df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2654df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 265553892102SStefano Zampini 2656df4d28bfSStefano Zampini ierr = ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2657df4d28bfSStefano Zampini ierr = ISDestroy(&reuse_solver->is_R);CHKERRQ(ierr); 265853892102SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->is_R_local);CHKERRQ(ierr); 2659df4d28bfSStefano Zampini reuse_solver->is_R = pcbddc->is_R_local; 266053892102SStefano Zampini } else { 26613a50541eSStefano Zampini ierr = PetscFree(idx_R_local);CHKERRQ(ierr); 266253892102SStefano Zampini } 2663a64d13efSStefano Zampini 2664a64d13efSStefano Zampini /* print some info if requested */ 2665a64d13efSStefano Zampini if (pcbddc->dbg_flag) { 2666a64d13efSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2667a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 26681575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2669a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);CHKERRQ(ierr); 2670a64d13efSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %d, dirichlet_size = %d, boundary_size = %d\n",pcis->n,n_D,n_B);CHKERRQ(ierr); 26714f1b2e48SStefano 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); 2672a64d13efSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2673a64d13efSStefano Zampini } 2674a64d13efSStefano Zampini 2675a64d13efSStefano Zampini /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */ 2676df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 26776816873aSStefano Zampini IS is_aux1,is_aux2; 26786816873aSStefano Zampini PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local; 26796816873aSStefano Zampini 26803a50541eSStefano Zampini ierr = ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2681854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);CHKERRQ(ierr); 2682854ce69bSBarry Smith ierr = PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);CHKERRQ(ierr); 2683a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 26844641a718SStefano Zampini for (i=0; i<n_D; i++) { 26854641a718SStefano Zampini ierr = PetscBTSet(bitmask,is_indices[i]);CHKERRQ(ierr); 26864641a718SStefano Zampini } 2687a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2688a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 26894641a718SStefano Zampini if (!PetscBTLookup(bitmask,idx_R_local[i])) { 26904641a718SStefano Zampini aux_array1[j++] = i; 2691a64d13efSStefano Zampini } 2692a64d13efSStefano Zampini } 2693a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2694a64d13efSStefano Zampini ierr = ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2695a64d13efSStefano Zampini for (i=0, j=0; i<n_B; i++) { 26964641a718SStefano Zampini if (!PetscBTLookup(bitmask,is_indices[i])) { 26974641a718SStefano Zampini aux_array2[j++] = i; 2698a64d13efSStefano Zampini } 2699a64d13efSStefano Zampini } 2700a64d13efSStefano Zampini ierr = ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);CHKERRQ(ierr); 2701a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);CHKERRQ(ierr); 2702a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);CHKERRQ(ierr); 2703a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2704a64d13efSStefano Zampini ierr = ISDestroy(&is_aux2);CHKERRQ(ierr); 2705a64d13efSStefano Zampini 27068eeda7d8SStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 2707785e854fSJed Brown ierr = PetscMalloc1(n_D,&aux_array1);CHKERRQ(ierr); 2708a64d13efSStefano Zampini for (i=0, j=0; i<n_R; i++) { 27094641a718SStefano Zampini if (PetscBTLookup(bitmask,idx_R_local[i])) { 27104641a718SStefano Zampini aux_array1[j++] = i; 2711a64d13efSStefano Zampini } 2712a64d13efSStefano Zampini } 2713a64d13efSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);CHKERRQ(ierr); 2714a64d13efSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 2715a64d13efSStefano Zampini ierr = ISDestroy(&is_aux1);CHKERRQ(ierr); 2716a64d13efSStefano Zampini } 27174641a718SStefano Zampini ierr = PetscBTDestroy(&bitmask);CHKERRQ(ierr); 27183a50541eSStefano Zampini ierr = ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);CHKERRQ(ierr); 2719d62866d3SStefano Zampini } else { 2720df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 27216816873aSStefano Zampini IS tis; 27226816873aSStefano Zampini PetscInt schur_size; 27236816873aSStefano Zampini 2724df4d28bfSStefano Zampini ierr = ISGetLocalSize(reuse_solver->is_B,&schur_size);CHKERRQ(ierr); 27256816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);CHKERRQ(ierr); 2726df4d28bfSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);CHKERRQ(ierr); 27276816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 27286816873aSStefano Zampini if (pcbddc->switch_static || pcbddc->dbg_flag) { 27296816873aSStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);CHKERRQ(ierr); 27306816873aSStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);CHKERRQ(ierr); 27316816873aSStefano Zampini ierr = ISDestroy(&tis);CHKERRQ(ierr); 2732d62866d3SStefano Zampini } 2733d62866d3SStefano Zampini } 2734a64d13efSStefano Zampini PetscFunctionReturn(0); 2735a64d13efSStefano Zampini } 2736a64d13efSStefano Zampini 2737304d26faSStefano Zampini 2738304d26faSStefano Zampini #undef __FUNCT__ 2739304d26faSStefano Zampini #define __FUNCT__ "PCBDDCSetUpLocalSolvers" 2740684f6988SStefano Zampini PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann) 2741304d26faSStefano Zampini { 2742304d26faSStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 2743304d26faSStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 2744304d26faSStefano Zampini PC pc_temp; 2745304d26faSStefano Zampini Mat A_RR; 2746f4ddd8eeSStefano Zampini MatReuse reuse; 2747304d26faSStefano Zampini PetscScalar m_one = -1.0; 2748304d26faSStefano Zampini PetscReal value; 274904708bb6SStefano Zampini PetscInt n_D,n_R; 27509577ea80SStefano Zampini PetscBool use_exact,use_exact_reduced,issbaij; 2751304d26faSStefano Zampini PetscErrorCode ierr; 2752e604994aSStefano Zampini /* prefixes stuff */ 2753312be037SStefano Zampini char dir_prefix[256],neu_prefix[256],str_level[16]; 2754e604994aSStefano Zampini size_t len; 2755304d26faSStefano Zampini 2756304d26faSStefano Zampini PetscFunctionBegin; 2757304d26faSStefano Zampini 2758e604994aSStefano Zampini /* compute prefixes */ 2759e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,"");CHKERRQ(ierr); 2760e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,"");CHKERRQ(ierr); 2761e604994aSStefano Zampini if (!pcbddc->current_level) { 2762e604994aSStefano Zampini ierr = PetscStrcpy(dir_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2763e604994aSStefano Zampini ierr = PetscStrcpy(neu_prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 2764e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2765e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2766e604994aSStefano Zampini } else { 2767e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 2768312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 2769e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 2770e604994aSStefano Zampini len -= 15; /* remove "pc_bddc_coarse_" */ 2771312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 2772312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 277334d6797cSStefano Zampini ierr = PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 277434d6797cSStefano Zampini ierr = PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 2775e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,"pc_bddc_dirichlet_");CHKERRQ(ierr); 2776e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,"pc_bddc_neumann_");CHKERRQ(ierr); 2777e604994aSStefano Zampini ierr = PetscStrcat(dir_prefix,str_level);CHKERRQ(ierr); 2778e604994aSStefano Zampini ierr = PetscStrcat(neu_prefix,str_level);CHKERRQ(ierr); 2779e604994aSStefano Zampini } 2780e604994aSStefano Zampini 2781304d26faSStefano Zampini /* DIRICHLET PROBLEM */ 2782684f6988SStefano Zampini if (dirichlet) { 2783d5574798SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2784450f8f5eSStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit) { 2785450f8f5eSStefano Zampini if (!sub_schurs->reuse_solver) { 2786450f8f5eSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented\n"); 2787450f8f5eSStefano Zampini } 2788450f8f5eSStefano Zampini if (pcbddc->dbg_flag) { 2789a3df083aSStefano Zampini Mat A_IIn; 2790a3df083aSStefano Zampini 2791a3df083aSStefano Zampini ierr = PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);CHKERRQ(ierr); 2792a3df083aSStefano Zampini ierr = MatDestroy(&pcis->A_II);CHKERRQ(ierr); 2793a3df083aSStefano Zampini pcis->A_II = A_IIn; 2794a3df083aSStefano Zampini } 2795450f8f5eSStefano Zampini } 27963301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 27973301b35fSStefano Zampini ierr = MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 2798964fefecSStefano Zampini } 2799ac78edfcSStefano Zampini /* Matrix for Dirichlet problem is pcis->A_II */ 2800964fefecSStefano Zampini n_D = pcis->n - pcis->n_B; 2801304d26faSStefano Zampini if (!pcbddc->ksp_D) { /* create object if not yet build */ 2802304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);CHKERRQ(ierr); 2803304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr); 2804304d26faSStefano Zampini /* default */ 2805304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_D,KSPPREONLY);CHKERRQ(ierr); 2806e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);CHKERRQ(ierr); 28079577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcis->A_II,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 2808304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 28099577ea80SStefano Zampini if (issbaij) { 28109577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 28119577ea80SStefano Zampini } else { 2812304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 28139577ea80SStefano Zampini } 2814304d26faSStefano Zampini /* Allow user's customization */ 2815304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_D);CHKERRQ(ierr); 2816304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2817304d26faSStefano Zampini } 2818d1e9a80fSBarry Smith ierr = KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr); 2819df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2820df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2821d62866d3SStefano Zampini 2822df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);CHKERRQ(ierr); 2823d5574798SStefano Zampini } 2824304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2825304d26faSStefano Zampini if (!n_D) { 2826304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_D,&pc_temp);CHKERRQ(ierr); 2827304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2828304d26faSStefano Zampini } 2829304d26faSStefano Zampini /* Set Up KSP for Dirichlet problem of BDDC */ 2830304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_D);CHKERRQ(ierr); 2831304d26faSStefano Zampini /* set ksp_D into pcis data */ 2832304d26faSStefano Zampini ierr = KSPDestroy(&pcis->ksp_D);CHKERRQ(ierr); 2833304d26faSStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ksp_D);CHKERRQ(ierr); 2834304d26faSStefano Zampini pcis->ksp_D = pcbddc->ksp_D; 2835684f6988SStefano Zampini } 2836304d26faSStefano Zampini 2837304d26faSStefano Zampini /* NEUMANN PROBLEM */ 2838684f6988SStefano Zampini A_RR = 0; 2839684f6988SStefano Zampini if (neumann) { 2840d62866d3SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 284104708bb6SStefano Zampini PetscInt ibs,mbs; 284204708bb6SStefano Zampini PetscBool issbaij; 284304708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 2844f4ddd8eeSStefano Zampini /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */ 28458ce42a96SStefano Zampini ierr = ISGetSize(pcbddc->is_R_local,&n_R);CHKERRQ(ierr); 2846f4ddd8eeSStefano Zampini if (pcbddc->ksp_R) { /* already created ksp */ 2847f4ddd8eeSStefano Zampini PetscInt nn_R; 284881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);CHKERRQ(ierr); 2849f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)A_RR);CHKERRQ(ierr); 2850f4ddd8eeSStefano Zampini ierr = MatGetSize(A_RR,&nn_R,NULL);CHKERRQ(ierr); 2851f4ddd8eeSStefano Zampini if (nn_R != n_R) { /* old ksp is not reusable, so reset it */ 2852f4ddd8eeSStefano Zampini ierr = KSPReset(pcbddc->ksp_R);CHKERRQ(ierr); 2853f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2854f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2855f4ddd8eeSStefano Zampini } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */ 2856727cdba6SStefano Zampini if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */ 2857f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2858f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2859f4ddd8eeSStefano Zampini } else { /* safe to reuse the matrix */ 2860f4ddd8eeSStefano Zampini reuse = MAT_REUSE_MATRIX; 2861f4ddd8eeSStefano Zampini } 2862f4ddd8eeSStefano Zampini } 2863f4ddd8eeSStefano Zampini /* last check */ 2864d1e9a80fSBarry Smith if (pc->flag == DIFFERENT_NONZERO_PATTERN) { 2865f4ddd8eeSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2866f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2867f4ddd8eeSStefano Zampini } 2868f4ddd8eeSStefano Zampini } else { /* first time, so we need to create the matrix */ 2869f4ddd8eeSStefano Zampini reuse = MAT_INITIAL_MATRIX; 2870f4ddd8eeSStefano Zampini } 2871f4ddd8eeSStefano Zampini /* extract A_RR */ 2872af732b37SStefano Zampini ierr = MatGetBlockSize(pcbddc->local_mat,&mbs);CHKERRQ(ierr); 2873af732b37SStefano Zampini ierr = ISGetBlockSize(pcbddc->is_R_local,&ibs);CHKERRQ(ierr); 287404708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 287504708bb6SStefano Zampini if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */ 287604708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 287704708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 287804708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 2879af732b37SStefano Zampini } else { 288004708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 28816816873aSStefano Zampini } 288204708bb6SStefano Zampini } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */ 288304708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 288404708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 288504708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 288604708bb6SStefano Zampini } else { 288704708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQBAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 288804708bb6SStefano Zampini } 288904708bb6SStefano Zampini } 289016e386b8SStefano Zampini if (pcbddc->benign_n && !pcbddc->benign_change_explicit && pcbddc->dbg_flag) { 289116e386b8SStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 289216e386b8SStefano Zampini ierr = PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);CHKERRQ(ierr); 289316e386b8SStefano Zampini } else { 2894f4ddd8eeSStefano Zampini ierr = MatGetSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);CHKERRQ(ierr); 289516e386b8SStefano Zampini } 28963301b35fSStefano Zampini if (pcbddc->local_mat->symmetric_set) { 28973301b35fSStefano Zampini ierr = MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric_set);CHKERRQ(ierr); 28986816873aSStefano Zampini } 2899f4ddd8eeSStefano Zampini if (!pcbddc->ksp_R) { /* create object if not present */ 2900304d26faSStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);CHKERRQ(ierr); 2901304d26faSStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);CHKERRQ(ierr); 2902304d26faSStefano Zampini /* default */ 2903304d26faSStefano Zampini ierr = KSPSetType(pcbddc->ksp_R,KSPPREONLY);CHKERRQ(ierr); 2904e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);CHKERRQ(ierr); 2905304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 29069577ea80SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr); 29079577ea80SStefano Zampini if (issbaij) { 29089577ea80SStefano Zampini ierr = PCSetType(pc_temp,PCCHOLESKY);CHKERRQ(ierr); 29099577ea80SStefano Zampini } else { 2910304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCLU);CHKERRQ(ierr); 29119577ea80SStefano Zampini } 2912304d26faSStefano Zampini /* Allow user's customization */ 2913304d26faSStefano Zampini ierr = KSPSetFromOptions(pcbddc->ksp_R);CHKERRQ(ierr); 2914304d26faSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 2915304d26faSStefano Zampini } 2916304d26faSStefano Zampini /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */ 2917304d26faSStefano Zampini if (!n_R) { 2918304d26faSStefano Zampini ierr = KSPGetPC(pcbddc->ksp_R,&pc_temp);CHKERRQ(ierr); 2919304d26faSStefano Zampini ierr = PCSetType(pc_temp,PCNONE);CHKERRQ(ierr); 2920304d26faSStefano Zampini } 2921df4d28bfSStefano Zampini /* Reuse solver if it is present */ 2922df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2923df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 2924d62866d3SStefano Zampini 2925df4d28bfSStefano Zampini ierr = KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);CHKERRQ(ierr); 2926d62866d3SStefano Zampini } 2927ffd830a3SStefano Zampini ierr = KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);CHKERRQ(ierr); 2928304d26faSStefano Zampini /* Set Up KSP for Neumann problem of BDDC */ 2929304d26faSStefano Zampini ierr = KSPSetUp(pcbddc->ksp_R);CHKERRQ(ierr); 2930684f6988SStefano Zampini } 29316816873aSStefano Zampini /* free Neumann problem's matrix */ 29326816873aSStefano Zampini ierr = MatDestroy(&A_RR);CHKERRQ(ierr); 2933304d26faSStefano Zampini 2934304d26faSStefano Zampini /* check Dirichlet and Neumann solvers and adapt them if a nullspace correction is needed */ 29350fccc4e9SStefano Zampini if (pcbddc->NullSpace || pcbddc->dbg_flag) { 2936684f6988SStefano Zampini if (pcbddc->dbg_flag) { 2937684f6988SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 29381575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 2939684f6988SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 2940684f6988SStefano Zampini } 2941684f6988SStefano Zampini if (dirichlet) { /* Dirichlet */ 29420fccc4e9SStefano Zampini ierr = VecSetRandom(pcis->vec1_D,NULL);CHKERRQ(ierr); 29430fccc4e9SStefano Zampini ierr = MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr); 29440fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);CHKERRQ(ierr); 29450fccc4e9SStefano Zampini ierr = VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);CHKERRQ(ierr); 29460fccc4e9SStefano Zampini ierr = VecNorm(pcis->vec1_D,NORM_INFINITY,&value);CHKERRQ(ierr); 2947304d26faSStefano Zampini /* need to be adapted? */ 2948b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2949b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2950b8ffe317SStefano Zampini ierr = PCBDDCSetUseExactDirichlet(pc,use_exact_reduced);CHKERRQ(ierr); 2951304d26faSStefano Zampini /* print info */ 2952304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2953e604994aSStefano 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); 2954304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2955304d26faSStefano Zampini } 2956b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced && !pcbddc->switch_static) { 2957298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcis->is_I_local);CHKERRQ(ierr); 2958304d26faSStefano Zampini } 2959684f6988SStefano Zampini } 2960684f6988SStefano Zampini if (neumann) { /* Neumann */ 29616816873aSStefano Zampini ierr = KSPGetOperators(pcbddc->ksp_R,&A_RR,NULL);CHKERRQ(ierr); 29620fccc4e9SStefano Zampini ierr = VecSetRandom(pcbddc->vec1_R,NULL);CHKERRQ(ierr); 29630fccc4e9SStefano Zampini ierr = MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);CHKERRQ(ierr); 29640fccc4e9SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);CHKERRQ(ierr); 29650fccc4e9SStefano Zampini ierr = VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);CHKERRQ(ierr); 29660fccc4e9SStefano Zampini ierr = VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);CHKERRQ(ierr); 2967304d26faSStefano Zampini /* need to be adapted? */ 2968b8ffe317SStefano Zampini use_exact = (PetscAbsReal(value) > 1.e-4 ? PETSC_FALSE : PETSC_TRUE); 2969b8ffe317SStefano Zampini ierr = MPI_Allreduce(&use_exact,&use_exact_reduced,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 2970304d26faSStefano Zampini /* print info */ 2971304d26faSStefano Zampini if (pcbddc->dbg_flag) { 2972e604994aSStefano 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); 2973304d26faSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 2974304d26faSStefano Zampini } 2975b8ffe317SStefano Zampini if (pcbddc->NullSpace && !use_exact_reduced) { /* is it the right logic? */ 2976298c0119SStefano Zampini ierr = PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->is_R_local);CHKERRQ(ierr); 2977304d26faSStefano Zampini } 29780fccc4e9SStefano Zampini } 2979684f6988SStefano Zampini } 2980304d26faSStefano Zampini PetscFunctionReturn(0); 2981304d26faSStefano Zampini } 2982304d26faSStefano Zampini 2983304d26faSStefano Zampini #undef __FUNCT__ 2984ba15a52eSStefano Zampini #define __FUNCT__ "PCBDDCSolveSubstructureCorrection" 298580677318SStefano Zampini static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose) 2986674ae819SStefano Zampini { 2987674ae819SStefano Zampini PetscErrorCode ierr; 2988674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 2989be83ff47SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 2990674ae819SStefano Zampini 2991674ae819SStefano Zampini PetscFunctionBegin; 2992df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 299380677318SStefano Zampini ierr = VecSet(pcbddc->vec1_R,0.);CHKERRQ(ierr); 299420c7b377SStefano Zampini } 299580677318SStefano Zampini if (!pcbddc->switch_static) { 299680677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 299780677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 299880677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 299920c7b377SStefano Zampini } 3000df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 300180677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 300280677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 300320c7b377SStefano Zampini } else { 3004df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3005be83ff47SStefano Zampini 3006df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3007df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 300820c7b377SStefano Zampini } 3009be83ff47SStefano Zampini } else { 301080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301480677318SStefano Zampini if (applytranspose && pcbddc->local_auxmat1) { 301580677318SStefano Zampini ierr = MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);CHKERRQ(ierr); 301680677318SStefano Zampini ierr = MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 301780677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 301880677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3019674ae819SStefano Zampini } 3020674ae819SStefano Zampini } 302172b8c272SStefano Zampini if (!sub_schurs->reuse_solver || pcbddc->switch_static) { 302280677318SStefano Zampini if (applytranspose) { 302380677318SStefano Zampini ierr = KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 302480677318SStefano Zampini } else { 302580677318SStefano Zampini ierr = KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 302680677318SStefano Zampini } 3027be83ff47SStefano Zampini } else { 3028df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3029be83ff47SStefano Zampini 3030be83ff47SStefano Zampini if (applytranspose) { 3031df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3032be83ff47SStefano Zampini } else { 3033df4d28bfSStefano Zampini ierr = MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);CHKERRQ(ierr); 3034be83ff47SStefano Zampini } 3035be83ff47SStefano Zampini } 303680677318SStefano Zampini ierr = VecSet(inout_B,0.);CHKERRQ(ierr); 303780677318SStefano Zampini if (!pcbddc->switch_static) { 3038df4d28bfSStefano Zampini if (!sub_schurs->reuse_solver) { 303980677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 304080677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3041be83ff47SStefano Zampini } else { 3042df4d28bfSStefano Zampini PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver; 3043be83ff47SStefano Zampini 3044df4d28bfSStefano Zampini ierr = VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3045df4d28bfSStefano Zampini ierr = VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3046be83ff47SStefano Zampini } 304780677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 304880677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 304980677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);CHKERRQ(ierr); 305080677318SStefano Zampini } 305180677318SStefano Zampini } else { 305280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305480677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305580677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 305680677318SStefano Zampini if (!applytranspose && pcbddc->local_auxmat1) { 305780677318SStefano Zampini ierr = MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);CHKERRQ(ierr); 305880677318SStefano Zampini ierr = MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);CHKERRQ(ierr); 305980677318SStefano Zampini } 306080677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306180677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306280677318SStefano Zampini ierr = VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 306380677318SStefano Zampini ierr = VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3064674ae819SStefano Zampini } 3065674ae819SStefano Zampini PetscFunctionReturn(0); 3066674ae819SStefano Zampini } 3067674ae819SStefano Zampini 3068dc359a40SStefano Zampini /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */ 3069674ae819SStefano Zampini #undef __FUNCT__ 3070674ae819SStefano Zampini #define __FUNCT__ "PCBDDCApplyInterfacePreconditioner" 3071dc359a40SStefano Zampini PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose) 3072674ae819SStefano Zampini { 3073674ae819SStefano Zampini PetscErrorCode ierr; 3074674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 3075674ae819SStefano Zampini PC_IS* pcis = (PC_IS*) (pc->data); 3076674ae819SStefano Zampini const PetscScalar zero = 0.0; 3077674ae819SStefano Zampini 3078674ae819SStefano Zampini PetscFunctionBegin; 3079dc359a40SStefano Zampini /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */ 30804fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3081dc359a40SStefano Zampini if (applytranspose) { 3082674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 30838eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 3084dc359a40SStefano Zampini } else { 3085674ae819SStefano Zampini ierr = MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);CHKERRQ(ierr); 3086674ae819SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P);CHKERRQ(ierr); } 308715aaf578SStefano Zampini } 30884fee134fSStefano Zampini } else { 30894fee134fSStefano Zampini ierr = VecSet(pcbddc->vec1_P,zero);CHKERRQ(ierr); 30904fee134fSStefano Zampini } 3091efc2fbd9SStefano Zampini 3092efc2fbd9SStefano Zampini /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */ 30934f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3094efc2fbd9SStefano Zampini PetscScalar *array; 30954f1b2e48SStefano Zampini PetscInt j; 3096efc2fbd9SStefano Zampini 3097efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 30984f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j]; 3099efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3100efc2fbd9SStefano Zampini } 3101efc2fbd9SStefano Zampini 310212edc857SStefano Zampini /* start communications from local primal nodes to rhs of coarse solver */ 310312edc857SStefano Zampini ierr = VecSet(pcbddc->coarse_vec,zero);CHKERRQ(ierr); 310412edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310512edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 310612edc857SStefano Zampini 31079f00e9b4SStefano Zampini /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */ 310812edc857SStefano Zampini /* TODO remove null space when doing multilevel */ 310912edc857SStefano Zampini if (pcbddc->coarse_ksp) { 311051694757SStefano Zampini Mat coarse_mat; 3111964fefecSStefano Zampini Vec rhs,sol; 311251694757SStefano Zampini MatNullSpace nullsp; 3113964fefecSStefano Zampini 3114964fefecSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&rhs);CHKERRQ(ierr); 3115964fefecSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&sol);CHKERRQ(ierr); 311651694757SStefano Zampini ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 311751694757SStefano Zampini ierr = MatGetNullSpace(coarse_mat,&nullsp);CHKERRQ(ierr); 311851694757SStefano Zampini if (nullsp) { 311951694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,rhs);CHKERRQ(ierr); 312051694757SStefano Zampini } 312112edc857SStefano Zampini if (applytranspose) { 3122964fefecSStefano Zampini ierr = KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 312312edc857SStefano Zampini } else { 3124964fefecSStefano Zampini ierr = KSPSolve(pcbddc->coarse_ksp,rhs,sol);CHKERRQ(ierr); 312512edc857SStefano Zampini } 312651694757SStefano Zampini if (nullsp) { 312751694757SStefano Zampini ierr = MatNullSpaceRemove(nullsp,sol);CHKERRQ(ierr); 312851694757SStefano Zampini } 312912edc857SStefano Zampini } 3130674ae819SStefano Zampini 3131674ae819SStefano Zampini /* Local solution on R nodes */ 31324fee134fSStefano Zampini if (pcis->n && !pcbddc->benign_apply_coarse_only) { 313380677318SStefano Zampini ierr = PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);CHKERRQ(ierr); 31349f00e9b4SStefano Zampini } 31359f00e9b4SStefano Zampini /* communications from coarse sol to local primal nodes */ 31369f00e9b4SStefano Zampini ierr = PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 313712edc857SStefano Zampini ierr = PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3138674ae819SStefano Zampini 31394fee134fSStefano Zampini /* Sum contributions from the two levels */ 31404fee134fSStefano Zampini if (!pcbddc->benign_apply_coarse_only) { 3141dc359a40SStefano Zampini if (applytranspose) { 3142dc359a40SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 3143dc359a40SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3144dc359a40SStefano Zampini } else { 3145674ae819SStefano Zampini ierr = MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 31468eeda7d8SStefano Zampini if (pcbddc->switch_static) { ierr = MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); } 3147dc359a40SStefano Zampini } 3148efc2fbd9SStefano Zampini /* store p0 */ 31494f1b2e48SStefano Zampini if (pcbddc->benign_n) { 3150efc2fbd9SStefano Zampini PetscScalar *array; 31514f1b2e48SStefano Zampini PetscInt j; 3152efc2fbd9SStefano Zampini 3153efc2fbd9SStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 31544f1b2e48SStefano Zampini for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j]; 3155efc2fbd9SStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 3156efc2fbd9SStefano Zampini } 31574fee134fSStefano Zampini } else { /* expand the coarse solution */ 31584fee134fSStefano Zampini if (applytranspose) { 31594fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31604fee134fSStefano Zampini } else { 31614fee134fSStefano Zampini ierr = MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);CHKERRQ(ierr); 31624fee134fSStefano Zampini } 31634fee134fSStefano Zampini } 3164674ae819SStefano Zampini PetscFunctionReturn(0); 3165674ae819SStefano Zampini } 3166674ae819SStefano Zampini 3167674ae819SStefano Zampini #undef __FUNCT__ 3168674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataBegin" 316912edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode) 3170674ae819SStefano Zampini { 3171674ae819SStefano Zampini PetscErrorCode ierr; 3172674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 317358da7f69SStefano Zampini PetscScalar *array; 317412edc857SStefano Zampini Vec from,to; 3175674ae819SStefano Zampini 3176674ae819SStefano Zampini PetscFunctionBegin; 317712edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 317812edc857SStefano Zampini from = pcbddc->coarse_vec; 317912edc857SStefano Zampini to = pcbddc->vec1_P; 318012edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 318112edc857SStefano Zampini Vec tvec; 318258da7f69SStefano Zampini 318358da7f69SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 318458da7f69SStefano Zampini ierr = VecResetArray(tvec);CHKERRQ(ierr); 318512edc857SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 318658da7f69SStefano Zampini ierr = VecGetArray(tvec,&array);CHKERRQ(ierr); 318758da7f69SStefano Zampini ierr = VecPlaceArray(from,array);CHKERRQ(ierr); 318858da7f69SStefano Zampini ierr = VecRestoreArray(tvec,&array);CHKERRQ(ierr); 318912edc857SStefano Zampini } 319012edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 319112edc857SStefano Zampini from = pcbddc->vec1_P; 319212edc857SStefano Zampini to = pcbddc->coarse_vec; 319312edc857SStefano Zampini } 319412edc857SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 3195674ae819SStefano Zampini PetscFunctionReturn(0); 3196674ae819SStefano Zampini } 3197674ae819SStefano Zampini 3198674ae819SStefano Zampini #undef __FUNCT__ 3199674ae819SStefano Zampini #define __FUNCT__ "PCBDDCScatterCoarseDataEnd" 320012edc857SStefano Zampini PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode) 3201674ae819SStefano Zampini { 3202674ae819SStefano Zampini PetscErrorCode ierr; 3203674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)(pc->data); 320458da7f69SStefano Zampini PetscScalar *array; 320512edc857SStefano Zampini Vec from,to; 3206674ae819SStefano Zampini 3207674ae819SStefano Zampini PetscFunctionBegin; 320812edc857SStefano Zampini if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */ 320912edc857SStefano Zampini from = pcbddc->coarse_vec; 321012edc857SStefano Zampini to = pcbddc->vec1_P; 321112edc857SStefano Zampini } else { /* from local to global -> put data in coarse right hand side */ 321212edc857SStefano Zampini from = pcbddc->vec1_P; 321312edc857SStefano Zampini to = pcbddc->coarse_vec; 321412edc857SStefano Zampini } 321512edc857SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);CHKERRQ(ierr); 321612edc857SStefano Zampini if (smode == SCATTER_FORWARD) { 321712edc857SStefano Zampini if (pcbddc->coarse_ksp) { /* get array from coarse processes */ 321812edc857SStefano Zampini Vec tvec; 321958da7f69SStefano Zampini 322012edc857SStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&tvec);CHKERRQ(ierr); 322158da7f69SStefano Zampini ierr = VecGetArray(to,&array);CHKERRQ(ierr); 322258da7f69SStefano Zampini ierr = VecPlaceArray(tvec,array);CHKERRQ(ierr); 322358da7f69SStefano Zampini ierr = VecRestoreArray(to,&array);CHKERRQ(ierr); 322458da7f69SStefano Zampini } 322558da7f69SStefano Zampini } else { 322658da7f69SStefano Zampini if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */ 322758da7f69SStefano Zampini ierr = VecResetArray(from);CHKERRQ(ierr); 322812edc857SStefano Zampini } 322912edc857SStefano Zampini } 3230674ae819SStefano Zampini PetscFunctionReturn(0); 3231674ae819SStefano Zampini } 3232674ae819SStefano Zampini 3233984c4197SStefano Zampini /* uncomment for testing purposes */ 3234984c4197SStefano Zampini /* #define PETSC_MISSING_LAPACK_GESVD 1 */ 3235674ae819SStefano Zampini #undef __FUNCT__ 3236674ae819SStefano Zampini #define __FUNCT__ "PCBDDCConstraintsSetUp" 3237674ae819SStefano Zampini PetscErrorCode PCBDDCConstraintsSetUp(PC pc) 3238674ae819SStefano Zampini { 3239674ae819SStefano Zampini PetscErrorCode ierr; 3240674ae819SStefano Zampini PC_IS* pcis = (PC_IS*)(pc->data); 3241674ae819SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 3242674ae819SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 3243984c4197SStefano Zampini /* one and zero */ 3244984c4197SStefano Zampini PetscScalar one=1.0,zero=0.0; 3245984c4197SStefano Zampini /* space to store constraints and their local indices */ 32469162d606SStefano Zampini PetscScalar *constraints_data; 32479162d606SStefano Zampini PetscInt *constraints_idxs,*constraints_idxs_B; 32489162d606SStefano Zampini PetscInt *constraints_idxs_ptr,*constraints_data_ptr; 32499162d606SStefano Zampini PetscInt *constraints_n; 3250984c4197SStefano Zampini /* iterators */ 3251b3d85658SStefano Zampini PetscInt i,j,k,total_counts,total_counts_cc,cum; 3252984c4197SStefano Zampini /* BLAS integers */ 3253e310c8b4SStefano Zampini PetscBLASInt lwork,lierr; 3254e310c8b4SStefano Zampini PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1; 3255c4303822SStefano Zampini PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC; 3256727cdba6SStefano Zampini /* reuse */ 32570e6343abSStefano Zampini PetscInt olocal_primal_size,olocal_primal_size_cc; 32580e6343abSStefano Zampini PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult; 3259984c4197SStefano Zampini /* change of basis */ 3260b3d85658SStefano Zampini PetscBool qr_needed; 32619162d606SStefano Zampini PetscBT change_basis,qr_needed_idx; 3262984c4197SStefano Zampini /* auxiliary stuff */ 326364efe560SStefano Zampini PetscInt *nnz,*is_indices; 32648a0068c3SStefano Zampini PetscInt ncc; 3265984c4197SStefano Zampini /* some quantities */ 326645a1bb75SStefano Zampini PetscInt n_vertices,total_primal_vertices,valid_constraints; 3267a58a30b4SStefano Zampini PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints; 3268984c4197SStefano Zampini 3269674ae819SStefano Zampini PetscFunctionBegin; 32708e61c736SStefano Zampini /* Destroy Mat objects computed previously */ 32718e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 32728e61c736SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 3273088faed8SStefano Zampini /* save info on constraints from previous setup (if any) */ 3274088faed8SStefano Zampini olocal_primal_size = pcbddc->local_primal_size; 32750e6343abSStefano Zampini olocal_primal_size_cc = pcbddc->local_primal_size_cc; 32760e6343abSStefano Zampini ierr = PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);CHKERRQ(ierr); 32770e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32780e6343abSStefano Zampini ierr = PetscMemcpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc*sizeof(PetscInt));CHKERRQ(ierr); 32790e6343abSStefano Zampini ierr = PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);CHKERRQ(ierr); 3280088faed8SStefano Zampini ierr = PetscFree(pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3281cf5a6209SStefano Zampini 3282cf5a6209SStefano Zampini /* print some info */ 3283cf5a6209SStefano Zampini if (pcbddc->dbg_flag) { 3284cf5a6209SStefano Zampini IS vertices; 3285cf5a6209SStefano Zampini PetscInt nv,nedges,nfaces; 3286cf5a6209SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&nfaces,NULL,&nedges,NULL,&vertices);CHKERRQ(ierr); 3287cf5a6209SStefano Zampini ierr = ISGetSize(vertices,&nv);CHKERRQ(ierr); 3288cf5a6209SStefano Zampini ierr = ISDestroy(&vertices);CHKERRQ(ierr); 32891575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3290cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3291cf5a6209SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%d)\n",PetscGlobalRank,nv,pcbddc->use_vertices);CHKERRQ(ierr); 3292fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%d)\n",PetscGlobalRank,nedges,pcbddc->use_edges);CHKERRQ(ierr); 3293fd14bc51SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%d)\n",PetscGlobalRank,nfaces,pcbddc->use_faces);CHKERRQ(ierr); 3294cf5a6209SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 32951575c14dSBarry Smith ierr = PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 3296cf5a6209SStefano Zampini } 3297cf5a6209SStefano Zampini 3298cf5a6209SStefano Zampini if (!pcbddc->adaptive_selection) { 32999162d606SStefano Zampini IS ISForVertices,*ISForFaces,*ISForEdges; 3300cf5a6209SStefano Zampini MatNullSpace nearnullsp; 3301cf5a6209SStefano Zampini const Vec *nearnullvecs; 3302cf5a6209SStefano Zampini Vec *localnearnullsp; 3303cf5a6209SStefano Zampini PetscScalar *array; 3304cf5a6209SStefano Zampini PetscInt n_ISForFaces,n_ISForEdges,nnsp_size; 3305cf5a6209SStefano Zampini PetscBool nnsp_has_cnst; 3306674ae819SStefano Zampini /* LAPACK working arrays for SVD or POD */ 3307b3d85658SStefano Zampini PetscBool skip_lapack,boolforchange; 3308674ae819SStefano Zampini PetscScalar *work; 3309674ae819SStefano Zampini PetscReal *singular_vals; 3310674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3311674ae819SStefano Zampini PetscReal *rwork; 3312674ae819SStefano Zampini #endif 3313674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3314674ae819SStefano Zampini PetscScalar *temp_basis,*correlation_mat; 3315674ae819SStefano Zampini #else 3316964fefecSStefano Zampini PetscBLASInt dummy_int=1; 3317964fefecSStefano Zampini PetscScalar dummy_scalar=1.; 3318674ae819SStefano Zampini #endif 3319674ae819SStefano Zampini 3320674ae819SStefano Zampini /* Get index sets for faces, edges and vertices from graph */ 3321d06fc5fdSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);CHKERRQ(ierr); 3322d06fc5fdSStefano Zampini /* free unneeded index sets */ 3323d06fc5fdSStefano Zampini if (!pcbddc->use_vertices) { 3324d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 3325674ae819SStefano Zampini } 3326d06fc5fdSStefano Zampini if (!pcbddc->use_edges) { 3327d06fc5fdSStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3328d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3329d06fc5fdSStefano Zampini } 3330d06fc5fdSStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3331d06fc5fdSStefano Zampini n_ISForEdges = 0; 3332d06fc5fdSStefano Zampini } 3333d06fc5fdSStefano Zampini if (!pcbddc->use_faces) { 3334d06fc5fdSStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3335d06fc5fdSStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3336d06fc5fdSStefano Zampini } 3337d06fc5fdSStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3338d06fc5fdSStefano Zampini n_ISForFaces = 0; 3339d06fc5fdSStefano Zampini } 334070022509SStefano Zampini 334170022509SStefano Zampini #if defined(PETSC_USE_DEBUG) 334270022509SStefano Zampini /* HACK: when solving singular problems not using vertices, a change of basis is mandatory. 334370022509SStefano Zampini Also use_change_of_basis should be consistent among processors */ 334470022509SStefano Zampini if (pcbddc->NullSpace) { 334570022509SStefano Zampini PetscBool tbool[2],gbool[2]; 334670022509SStefano Zampini 334770022509SStefano Zampini if (!ISForVertices && !pcbddc->user_ChangeOfBasisMatrix) { 3348b8ffe317SStefano Zampini pcbddc->use_change_of_basis = PETSC_TRUE; 3349d06fc5fdSStefano Zampini if (!ISForEdges) { 3350d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = PETSC_TRUE; 3351d06fc5fdSStefano Zampini } 3352b8ffe317SStefano Zampini } 3353d06fc5fdSStefano Zampini tbool[0] = pcbddc->use_change_of_basis; 3354d06fc5fdSStefano Zampini tbool[1] = pcbddc->use_change_on_faces; 3355d06fc5fdSStefano Zampini ierr = MPI_Allreduce(tbool,gbool,2,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 3356d06fc5fdSStefano Zampini pcbddc->use_change_of_basis = gbool[0]; 3357d06fc5fdSStefano Zampini pcbddc->use_change_on_faces = gbool[1]; 335898a51de6SStefano Zampini } 335970022509SStefano Zampini #endif 336008122e43SStefano Zampini 3361674ae819SStefano Zampini /* check if near null space is attached to global mat */ 3362674ae819SStefano Zampini ierr = MatGetNearNullSpace(pc->pmat,&nearnullsp);CHKERRQ(ierr); 3363674ae819SStefano Zampini if (nearnullsp) { 3364674ae819SStefano Zampini ierr = MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);CHKERRQ(ierr); 3365f4ddd8eeSStefano Zampini /* remove any stored info */ 3366f4ddd8eeSStefano Zampini ierr = MatNullSpaceDestroy(&pcbddc->onearnullspace);CHKERRQ(ierr); 3367f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3368f4ddd8eeSStefano Zampini /* store information for BDDC solver reuse */ 3369f4ddd8eeSStefano Zampini ierr = PetscObjectReference((PetscObject)nearnullsp);CHKERRQ(ierr); 3370f4ddd8eeSStefano Zampini pcbddc->onearnullspace = nearnullsp; 3371473ba861SJed Brown ierr = PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);CHKERRQ(ierr); 3372f4ddd8eeSStefano Zampini for (i=0;i<nnsp_size;i++) { 3373f4ddd8eeSStefano Zampini ierr = PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);CHKERRQ(ierr); 3374f4ddd8eeSStefano Zampini } 3375984c4197SStefano Zampini } else { /* if near null space is not provided BDDC uses constants by default */ 3376984c4197SStefano Zampini nnsp_size = 0; 3377674ae819SStefano Zampini nnsp_has_cnst = PETSC_TRUE; 3378674ae819SStefano Zampini } 3379984c4197SStefano Zampini /* get max number of constraints on a single cc */ 3380984c4197SStefano Zampini max_constraints = nnsp_size; 3381984c4197SStefano Zampini if (nnsp_has_cnst) max_constraints++; 3382984c4197SStefano Zampini 3383674ae819SStefano Zampini /* 3384674ae819SStefano Zampini Evaluate maximum storage size needed by the procedure 33859162d606SStefano Zampini - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]" 33869162d606SStefano Zampini - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]" 33879162d606SStefano Zampini There can be multiple constraints per connected component 3388674ae819SStefano Zampini */ 3389674ae819SStefano Zampini n_vertices = 0; 3390674ae819SStefano Zampini if (ISForVertices) { 3391674ae819SStefano Zampini ierr = ISGetSize(ISForVertices,&n_vertices);CHKERRQ(ierr); 3392674ae819SStefano Zampini } 33939162d606SStefano Zampini ncc = n_vertices+n_ISForFaces+n_ISForEdges; 33949162d606SStefano Zampini ierr = PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);CHKERRQ(ierr); 33959162d606SStefano Zampini 33969162d606SStefano Zampini total_counts = n_ISForFaces+n_ISForEdges; 33979162d606SStefano Zampini total_counts *= max_constraints; 3398674ae819SStefano Zampini total_counts += n_vertices; 33994641a718SStefano Zampini ierr = PetscBTCreate(total_counts,&change_basis);CHKERRQ(ierr); 34009162d606SStefano Zampini 3401674ae819SStefano Zampini total_counts = 0; 3402674ae819SStefano Zampini max_size_of_constraint = 0; 3403674ae819SStefano Zampini for (i=0;i<n_ISForEdges+n_ISForFaces;i++) { 34049162d606SStefano Zampini IS used_is; 3405674ae819SStefano Zampini if (i<n_ISForEdges) { 34069162d606SStefano Zampini used_is = ISForEdges[i]; 3407674ae819SStefano Zampini } else { 34089162d606SStefano Zampini used_is = ISForFaces[i-n_ISForEdges]; 3409674ae819SStefano Zampini } 34109162d606SStefano Zampini ierr = ISGetSize(used_is,&j);CHKERRQ(ierr); 3411674ae819SStefano Zampini total_counts += j; 3412674ae819SStefano Zampini max_size_of_constraint = PetscMax(j,max_size_of_constraint); 3413674ae819SStefano Zampini } 34149162d606SStefano 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); 34159162d606SStefano Zampini 3416984c4197SStefano Zampini /* get local part of global near null space vectors */ 3417785e854fSJed Brown ierr = PetscMalloc1(nnsp_size,&localnearnullsp);CHKERRQ(ierr); 3418984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3419984c4197SStefano Zampini ierr = VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);CHKERRQ(ierr); 3420e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3421e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3422984c4197SStefano Zampini } 3423674ae819SStefano Zampini 3424242a89d7SStefano Zampini /* whether or not to skip lapack calls */ 3425242a89d7SStefano Zampini skip_lapack = PETSC_TRUE; 3426a773dcb8SStefano Zampini if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE; 3427242a89d7SStefano Zampini 3428984c4197SStefano Zampini /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */ 3429a773dcb8SStefano Zampini if (!skip_lapack) { 3430674ae819SStefano Zampini PetscScalar temp_work; 3431911cabfeSStefano Zampini 3432674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3433984c4197SStefano Zampini /* Proper Orthogonal Decomposition (POD) using the snapshot method */ 3434785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&correlation_mat);CHKERRQ(ierr); 3435785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&singular_vals);CHKERRQ(ierr); 3436785e854fSJed Brown ierr = PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);CHKERRQ(ierr); 3437674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3438785e854fSJed Brown ierr = PetscMalloc1(3*max_constraints,&rwork);CHKERRQ(ierr); 3439674ae819SStefano Zampini #endif 3440674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3441c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 3442c8244a33SStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_LDA);CHKERRQ(ierr); 3443674ae819SStefano Zampini lwork = -1; 3444674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3445674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3446c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr)); 3447674ae819SStefano Zampini #else 3448c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr)); 3449674ae819SStefano Zampini #endif 3450674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3451984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr); 3452674ae819SStefano Zampini #else /* on missing GESVD */ 3453674ae819SStefano Zampini /* SVD */ 3454674ae819SStefano Zampini PetscInt max_n,min_n; 3455674ae819SStefano Zampini max_n = max_size_of_constraint; 3456984c4197SStefano Zampini min_n = max_constraints; 3457984c4197SStefano Zampini if (max_size_of_constraint < max_constraints) { 3458674ae819SStefano Zampini min_n = max_size_of_constraint; 3459984c4197SStefano Zampini max_n = max_constraints; 3460674ae819SStefano Zampini } 3461785e854fSJed Brown ierr = PetscMalloc1(min_n,&singular_vals);CHKERRQ(ierr); 3462674ae819SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3463785e854fSJed Brown ierr = PetscMalloc1(5*min_n,&rwork);CHKERRQ(ierr); 3464674ae819SStefano Zampini #endif 3465674ae819SStefano Zampini /* now we evaluate the optimal workspace using query with lwork=-1 */ 3466674ae819SStefano Zampini lwork = -1; 3467e310c8b4SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_M);CHKERRQ(ierr); 3468e310c8b4SStefano Zampini ierr = PetscBLASIntCast(min_n,&Blas_N);CHKERRQ(ierr); 3469b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(max_n,&Blas_LDA);CHKERRQ(ierr); 3470674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3471674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 34729162d606SStefano 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)); 3473674ae819SStefano Zampini #else 34749162d606SStefano 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)); 3475674ae819SStefano Zampini #endif 3476674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3477984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr); 3478984c4197SStefano Zampini #endif /* on missing GESVD */ 3479674ae819SStefano Zampini /* Allocate optimal workspace */ 3480674ae819SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);CHKERRQ(ierr); 3481854ce69bSBarry Smith ierr = PetscMalloc1(lwork,&work);CHKERRQ(ierr); 3482674ae819SStefano Zampini } 3483674ae819SStefano Zampini /* Now we can loop on constraining sets */ 3484674ae819SStefano Zampini total_counts = 0; 34859162d606SStefano Zampini constraints_idxs_ptr[0] = 0; 34869162d606SStefano Zampini constraints_data_ptr[0] = 0; 3487674ae819SStefano Zampini /* vertices */ 34889162d606SStefano Zampini if (n_vertices) { 3489674ae819SStefano Zampini ierr = ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 34909162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs,is_indices,n_vertices*sizeof(PetscInt));CHKERRQ(ierr); 3491674ae819SStefano Zampini for (i=0;i<n_vertices;i++) { 34929162d606SStefano Zampini constraints_n[total_counts] = 1; 34939162d606SStefano Zampini constraints_data[total_counts] = 1.0; 34949162d606SStefano Zampini constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1; 34959162d606SStefano Zampini constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1; 3496674ae819SStefano Zampini total_counts++; 3497674ae819SStefano Zampini } 3498674ae819SStefano Zampini ierr = ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3499674ae819SStefano Zampini n_vertices = total_counts; 3500674ae819SStefano Zampini } 3501984c4197SStefano Zampini 3502674ae819SStefano Zampini /* edges and faces */ 35039162d606SStefano Zampini total_counts_cc = total_counts; 3504911cabfeSStefano Zampini for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) { 35059162d606SStefano Zampini IS used_is; 35069162d606SStefano Zampini PetscBool idxs_copied = PETSC_FALSE; 35079162d606SStefano Zampini 3508911cabfeSStefano Zampini if (ncc<n_ISForEdges) { 35099162d606SStefano Zampini used_is = ISForEdges[ncc]; 3510984c4197SStefano Zampini boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */ 3511674ae819SStefano Zampini } else { 35129162d606SStefano Zampini used_is = ISForFaces[ncc-n_ISForEdges]; 3513984c4197SStefano Zampini boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */ 3514674ae819SStefano Zampini } 3515674ae819SStefano Zampini temp_constraints = 0; /* zero the number of constraints I have on this conn comp */ 35169162d606SStefano Zampini 35179162d606SStefano Zampini ierr = ISGetSize(used_is,&size_of_constraint);CHKERRQ(ierr); 35189162d606SStefano Zampini ierr = ISGetIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 3519984c4197SStefano Zampini /* change of basis should not be performed on local periodic nodes */ 3520984c4197SStefano Zampini if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE; 3521674ae819SStefano Zampini if (nnsp_has_cnst) { 35225b08dc53SStefano Zampini PetscScalar quad_value; 35239162d606SStefano Zampini 35249162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35259162d606SStefano Zampini idxs_copied = PETSC_TRUE; 35269162d606SStefano Zampini 3527a773dcb8SStefano Zampini if (!pcbddc->use_nnsp_true) { 3528674ae819SStefano Zampini quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint)); 3529a773dcb8SStefano Zampini } else { 3530a773dcb8SStefano Zampini quad_value = 1.0; 3531a773dcb8SStefano Zampini } 3532674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35339162d606SStefano Zampini constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value; 3534674ae819SStefano Zampini } 35359162d606SStefano Zampini temp_constraints++; 3536674ae819SStefano Zampini total_counts++; 3537674ae819SStefano Zampini } 3538674ae819SStefano Zampini for (k=0;k<nnsp_size;k++) { 3539984c4197SStefano Zampini PetscReal real_value; 35409162d606SStefano Zampini PetscScalar *ptr_to_data; 35419162d606SStefano Zampini 3542984c4197SStefano Zampini ierr = VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 35439162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint]; 3544674ae819SStefano Zampini for (j=0;j<size_of_constraint;j++) { 35459162d606SStefano Zampini ptr_to_data[j] = array[is_indices[j]]; 3546674ae819SStefano Zampini } 3547984c4197SStefano Zampini ierr = VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);CHKERRQ(ierr); 3548984c4197SStefano Zampini /* check if array is null on the connected component */ 3549e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35509162d606SStefano Zampini PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one)); 35515b08dc53SStefano Zampini if (real_value > 0.0) { /* keep indices and values */ 3552674ae819SStefano Zampini temp_constraints++; 3553674ae819SStefano Zampini total_counts++; 35549162d606SStefano Zampini if (!idxs_copied) { 35559162d606SStefano Zampini ierr = PetscMemcpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint*sizeof(PetscInt));CHKERRQ(ierr); 35569162d606SStefano Zampini idxs_copied = PETSC_TRUE; 3557674ae819SStefano Zampini } 3558674ae819SStefano Zampini } 35599162d606SStefano Zampini } 35609162d606SStefano Zampini ierr = ISRestoreIndices(used_is,(const PetscInt**)&is_indices);CHKERRQ(ierr); 356145a1bb75SStefano Zampini valid_constraints = temp_constraints; 3562eb97c9d2SStefano Zampini if (!pcbddc->use_nnsp_true && temp_constraints) { 3563a773dcb8SStefano Zampini if (temp_constraints == 1) { /* just normalize the constraint */ 35649162d606SStefano Zampini PetscScalar norm,*ptr_to_data; 35659162d606SStefano Zampini 35669162d606SStefano Zampini ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3567a773dcb8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 35689162d606SStefano Zampini PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one)); 3569a773dcb8SStefano Zampini norm = 1.0/PetscSqrtReal(PetscRealPart(norm)); 35709162d606SStefano Zampini PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one)); 3571a773dcb8SStefano Zampini } else { /* perform SVD */ 3572984c4197SStefano Zampini PetscReal tol = 1.0e-8; /* tolerance for retaining eigenmodes */ 35739162d606SStefano Zampini PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]]; 3574674ae819SStefano Zampini 3575674ae819SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3576984c4197SStefano Zampini /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag 3577984c4197SStefano Zampini POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2) 3578984c4197SStefano Zampini -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined 3579984c4197SStefano Zampini the constraints basis will differ (by a complex factor with absolute value equal to 1) 3580984c4197SStefano Zampini from that computed using LAPACKgesvd 3581984c4197SStefano Zampini -> This is due to a different computation of eigenvectors in LAPACKheev 3582984c4197SStefano Zampini -> The quality of the POD-computed basis will be the same */ 3583984c4197SStefano Zampini ierr = PetscMemzero(correlation_mat,temp_constraints*temp_constraints*sizeof(PetscScalar));CHKERRQ(ierr); 3584674ae819SStefano Zampini /* Store upper triangular part of correlation matrix */ 3585e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 3586984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3587674ae819SStefano Zampini for (j=0;j<temp_constraints;j++) { 3588674ae819SStefano Zampini for (k=0;k<j+1;k++) { 35899162d606SStefano 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)); 3590674ae819SStefano Zampini } 3591674ae819SStefano Zampini } 3592e310c8b4SStefano Zampini /* compute eigenvalues and eigenvectors of correlation matrix */ 3593e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3594e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDA);CHKERRQ(ierr); 3595674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3596c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr)); 3597674ae819SStefano Zampini #else 3598c8244a33SStefano Zampini PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr)); 3599674ae819SStefano Zampini #endif 3600674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3601984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr); 3602984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */ 3603674ae819SStefano Zampini j = 0; 3604984c4197SStefano Zampini while (j < temp_constraints && singular_vals[j] < tol) j++; 3605674ae819SStefano Zampini total_counts = total_counts-j; 360645a1bb75SStefano Zampini valid_constraints = temp_constraints-j; 3607e310c8b4SStefano Zampini /* scale and copy POD basis into used quadrature memory */ 3608c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3609c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3610c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_K);CHKERRQ(ierr); 3611c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3612c4303822SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_LDB);CHKERRQ(ierr); 3613c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 3614674ae819SStefano Zampini if (j<temp_constraints) { 3615984c4197SStefano Zampini PetscInt ii; 3616984c4197SStefano Zampini for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]); 3617674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 36189162d606SStefano 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)); 3619674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3620984c4197SStefano Zampini for (k=0;k<temp_constraints-j;k++) { 3621674ae819SStefano Zampini for (ii=0;ii<size_of_constraint;ii++) { 36229162d606SStefano 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]; 3623674ae819SStefano Zampini } 3624674ae819SStefano Zampini } 3625674ae819SStefano Zampini } 3626674ae819SStefano Zampini #else /* on missing GESVD */ 3627e310c8b4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 3628e310c8b4SStefano Zampini ierr = PetscBLASIntCast(temp_constraints,&Blas_N);CHKERRQ(ierr); 3629b7d8b9f8SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3630674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 3631674ae819SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 36329162d606SStefano 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)); 3633674ae819SStefano Zampini #else 36349162d606SStefano 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)); 3635674ae819SStefano Zampini #endif 3636984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr); 3637674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3638984c4197SStefano Zampini /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */ 3639e310c8b4SStefano Zampini k = temp_constraints; 3640e310c8b4SStefano Zampini if (k > size_of_constraint) k = size_of_constraint; 3641674ae819SStefano Zampini j = 0; 3642e310c8b4SStefano Zampini while (j < k && singular_vals[k-j-1] < tol) j++; 364345a1bb75SStefano Zampini valid_constraints = k-j; 3644911cabfeSStefano Zampini total_counts = total_counts-temp_constraints+valid_constraints; 3645984c4197SStefano Zampini #endif /* on missing GESVD */ 3646674ae819SStefano Zampini } 3647a773dcb8SStefano Zampini } 36489162d606SStefano Zampini /* update pointers information */ 36499162d606SStefano Zampini if (valid_constraints) { 36509162d606SStefano Zampini constraints_n[total_counts_cc] = valid_constraints; 36519162d606SStefano Zampini constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint; 36529162d606SStefano Zampini constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints; 36539162d606SStefano Zampini /* set change_of_basis flag */ 365445a1bb75SStefano Zampini if (boolforchange) { 3655b3d85658SStefano Zampini PetscBTSet(change_basis,total_counts_cc); 36569162d606SStefano Zampini } 3657b3d85658SStefano Zampini total_counts_cc++; 365845a1bb75SStefano Zampini } 365945a1bb75SStefano Zampini } 3660984c4197SStefano Zampini /* free workspace */ 36618f1c130eSStefano Zampini if (!skip_lapack) { 3662984c4197SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 3663984c4197SStefano Zampini #if defined(PETSC_USE_COMPLEX) 3664984c4197SStefano Zampini ierr = PetscFree(rwork);CHKERRQ(ierr); 3665984c4197SStefano Zampini #endif 3666984c4197SStefano Zampini ierr = PetscFree(singular_vals);CHKERRQ(ierr); 3667984c4197SStefano Zampini #if defined(PETSC_MISSING_LAPACK_GESVD) 3668984c4197SStefano Zampini ierr = PetscFree(correlation_mat);CHKERRQ(ierr); 3669984c4197SStefano Zampini ierr = PetscFree(temp_basis);CHKERRQ(ierr); 3670984c4197SStefano Zampini #endif 3671984c4197SStefano Zampini } 3672984c4197SStefano Zampini for (k=0;k<nnsp_size;k++) { 3673984c4197SStefano Zampini ierr = VecDestroy(&localnearnullsp[k]);CHKERRQ(ierr); 3674984c4197SStefano Zampini } 3675984c4197SStefano Zampini ierr = PetscFree(localnearnullsp);CHKERRQ(ierr); 3676cf5a6209SStefano Zampini /* free index sets of faces, edges and vertices */ 3677cf5a6209SStefano Zampini for (i=0;i<n_ISForFaces;i++) { 3678cf5a6209SStefano Zampini ierr = ISDestroy(&ISForFaces[i]);CHKERRQ(ierr); 3679cf5a6209SStefano Zampini } 3680cf5a6209SStefano Zampini if (n_ISForFaces) { 3681cf5a6209SStefano Zampini ierr = PetscFree(ISForFaces);CHKERRQ(ierr); 3682cf5a6209SStefano Zampini } 3683cf5a6209SStefano Zampini for (i=0;i<n_ISForEdges;i++) { 3684cf5a6209SStefano Zampini ierr = ISDestroy(&ISForEdges[i]);CHKERRQ(ierr); 3685cf5a6209SStefano Zampini } 3686cf5a6209SStefano Zampini if (n_ISForEdges) { 3687cf5a6209SStefano Zampini ierr = PetscFree(ISForEdges);CHKERRQ(ierr); 3688cf5a6209SStefano Zampini } 3689cf5a6209SStefano Zampini ierr = ISDestroy(&ISForVertices);CHKERRQ(ierr); 369008122e43SStefano Zampini } else { 369108122e43SStefano Zampini PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 3692984c4197SStefano Zampini 369308122e43SStefano Zampini total_counts = 0; 369408122e43SStefano Zampini n_vertices = 0; 3695d62866d3SStefano Zampini if (sub_schurs->is_vertices && pcbddc->use_vertices) { 3696d62866d3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 369708122e43SStefano Zampini } 369808122e43SStefano Zampini max_constraints = 0; 36999162d606SStefano Zampini total_counts_cc = 0; 370008122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 370108122e43SStefano Zampini total_counts += pcbddc->adaptive_constraints_n[i]; 37029162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++; 370308122e43SStefano Zampini max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]); 370408122e43SStefano Zampini } 37059162d606SStefano Zampini constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr; 37069162d606SStefano Zampini constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr; 37079162d606SStefano Zampini constraints_idxs = pcbddc->adaptive_constraints_idxs; 37089162d606SStefano Zampini constraints_data = pcbddc->adaptive_constraints_data; 370974d5cdf7SStefano Zampini /* constraints_n differs from pcbddc->adaptive_constraints_n */ 37109162d606SStefano Zampini ierr = PetscMalloc1(total_counts_cc,&constraints_n);CHKERRQ(ierr); 37119162d606SStefano Zampini total_counts_cc = 0; 37129162d606SStefano Zampini for (i=0;i<sub_schurs->n_subs+n_vertices;i++) { 37139162d606SStefano Zampini if (pcbddc->adaptive_constraints_n[i]) { 37149162d606SStefano Zampini constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i]; 371508122e43SStefano Zampini } 371608122e43SStefano Zampini } 37179162d606SStefano Zampini #if 0 37189162d606SStefano Zampini printf("Found %d totals (%d)\n",total_counts_cc,total_counts); 37199162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37209162d606SStefano Zampini printf("const %d, start %d",i,constraints_idxs_ptr[i]); 37219162d606SStefano Zampini printf(" end %d:\n",constraints_idxs_ptr[i+1]); 37229162d606SStefano Zampini for (j=constraints_idxs_ptr[i];j<constraints_idxs_ptr[i+1];j++) { 37239162d606SStefano Zampini printf(" %d",constraints_idxs[j]); 37249162d606SStefano Zampini } 37259162d606SStefano Zampini printf("\n"); 37269162d606SStefano Zampini printf("number of cc: %d\n",constraints_n[i]); 37279162d606SStefano Zampini } 37281b968477SStefano Zampini for (i=0;i<n_vertices;i++) { 37298bec7fa6SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] vertex %d, n %d\n",PetscGlobalRank,i,pcbddc->adaptive_constraints_n[i]); 37301b968477SStefano Zampini } 37311b968477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 37328bec7fa6SStefano 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]); 37331b968477SStefano Zampini } 373408122e43SStefano Zampini #endif 373508122e43SStefano Zampini 37368bec7fa6SStefano Zampini max_size_of_constraint = 0; 37379162d606SStefano 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]); 37389162d606SStefano Zampini ierr = PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);CHKERRQ(ierr); 373908122e43SStefano Zampini /* Change of basis */ 3740b3d85658SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&change_basis);CHKERRQ(ierr); 374108122e43SStefano Zampini if (pcbddc->use_change_of_basis) { 374208122e43SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 374308122e43SStefano Zampini if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) { 3744b3d85658SStefano Zampini ierr = PetscBTSet(change_basis,i+n_vertices);CHKERRQ(ierr); 374508122e43SStefano Zampini } 374608122e43SStefano Zampini } 374708122e43SStefano Zampini } 374808122e43SStefano Zampini } 3749984c4197SStefano Zampini pcbddc->local_primal_size = total_counts; 37504f1b2e48SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 375108122e43SStefano Zampini 37529162d606SStefano Zampini /* map constraints_idxs in boundary numbering */ 37539162d606SStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);CHKERRQ(ierr); 37549162d606SStefano Zampini if (i != constraints_idxs_ptr[total_counts_cc]) { 37559162d606SStefano 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); 375608122e43SStefano Zampini } 3757674ae819SStefano Zampini 3758674ae819SStefano Zampini /* Create constraint matrix */ 3759674ae819SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 376016f15bc4SStefano Zampini ierr = MatSetType(pcbddc->ConstraintMatrix,MATAIJ);CHKERRQ(ierr); 3761984c4197SStefano Zampini ierr = MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);CHKERRQ(ierr); 3762984c4197SStefano Zampini 3763984c4197SStefano Zampini /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */ 3764a717540cSStefano Zampini /* determine if a QR strategy is needed for change of basis */ 3765a717540cSStefano Zampini qr_needed = PETSC_FALSE; 376674d5cdf7SStefano Zampini ierr = PetscBTCreate(total_counts_cc,&qr_needed_idx);CHKERRQ(ierr); 3767984c4197SStefano Zampini total_primal_vertices=0; 3768b3d85658SStefano Zampini pcbddc->local_primal_size_cc = 0; 37699162d606SStefano Zampini for (i=0;i<total_counts_cc;i++) { 37709162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 377172b8c272SStefano Zampini if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) { 37729162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]]; 3773b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 377464efe560SStefano Zampini } else if (PetscBTLookup(change_basis,i)) { 37759162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 37769162d606SStefano Zampini pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 3777a717540cSStefano Zampini } 3778b3d85658SStefano Zampini pcbddc->local_primal_size_cc += constraints_n[i]; 377974d5cdf7SStefano Zampini if (constraints_n[i] > 1 || pcbddc->use_qr_single || pcbddc->faster_deluxe) { 3780a717540cSStefano Zampini PetscBTSet(qr_needed_idx,i); 3781a717540cSStefano Zampini qr_needed = PETSC_TRUE; 3782a717540cSStefano Zampini } 3783fa434743SStefano Zampini } else { 3784b3d85658SStefano Zampini pcbddc->local_primal_size_cc += 1; 3785fa434743SStefano Zampini } 3786a717540cSStefano Zampini } 3787b371cd4fSStefano Zampini /* note that the local variable n_vertices used below stores the number of pointwise constraints */ 3788b371cd4fSStefano Zampini pcbddc->n_vertices = total_primal_vertices; 3789674ae819SStefano Zampini /* permute indices in order to have a sorted set of vertices */ 379070022509SStefano Zampini ierr = PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);CHKERRQ(ierr); 3791b3d85658SStefano Zampini 37924f1b2e48SStefano 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); 37930e6343abSStefano Zampini ierr = PetscMemcpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices*sizeof(PetscInt));CHKERRQ(ierr); 37940e6343abSStefano Zampini for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1; 3795984c4197SStefano Zampini 3796984c4197SStefano Zampini /* nonzero structure of constraint matrix */ 379774d5cdf7SStefano Zampini /* and get reference dof for local constraints */ 3798785e854fSJed Brown ierr = PetscMalloc1(pcbddc->local_primal_size,&nnz);CHKERRQ(ierr); 3799984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) nnz[i] = 1; 380074d5cdf7SStefano Zampini 3801984c4197SStefano Zampini j = total_primal_vertices; 380274d5cdf7SStefano Zampini total_counts = total_primal_vertices; 3803b3d85658SStefano Zampini cum = total_primal_vertices; 38049162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38054641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 3806b3d85658SStefano Zampini pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]]; 3807b3d85658SStefano Zampini pcbddc->local_primal_ref_mult[cum] = constraints_n[i]; 3808b3d85658SStefano Zampini cum++; 38099162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 381074d5cdf7SStefano Zampini for (k=0;k<constraints_n[i];k++) { 381174d5cdf7SStefano Zampini pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k]; 381274d5cdf7SStefano Zampini nnz[j+k] = size_of_constraint; 381374d5cdf7SStefano Zampini } 38149162d606SStefano Zampini j += constraints_n[i]; 3815674ae819SStefano Zampini } 3816674ae819SStefano Zampini } 3817674ae819SStefano Zampini ierr = MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);CHKERRQ(ierr); 3818674ae819SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3819088faed8SStefano Zampini 3820674ae819SStefano Zampini /* set values in constraint matrix */ 3821984c4197SStefano Zampini for (i=0;i<total_primal_vertices;i++) { 38220e6343abSStefano Zampini ierr = MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 3823674ae819SStefano Zampini } 3824984c4197SStefano Zampini total_counts = total_primal_vertices; 38259162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 38264641a718SStefano Zampini if (!PetscBTLookup(change_basis,i)) { 38279162d606SStefano Zampini PetscInt *cols; 38289162d606SStefano Zampini 38299162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 38309162d606SStefano Zampini cols = constraints_idxs+constraints_idxs_ptr[i]; 38319162d606SStefano Zampini for (k=0;k<constraints_n[i];k++) { 38329162d606SStefano Zampini PetscInt row = total_counts+k; 38339162d606SStefano Zampini PetscScalar *vals; 38349162d606SStefano Zampini 38359162d606SStefano Zampini vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint; 38369162d606SStefano Zampini ierr = MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 38379162d606SStefano Zampini } 38389162d606SStefano Zampini total_counts += constraints_n[i]; 3839674ae819SStefano Zampini } 3840674ae819SStefano Zampini } 3841674ae819SStefano Zampini /* assembling */ 3842674ae819SStefano Zampini ierr = MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3843674ae819SStefano Zampini ierr = MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3844088faed8SStefano Zampini 3845984c4197SStefano Zampini /* 384645a1bb75SStefano Zampini ierr = PetscViewerSetFormat(PETSC_VIEWER_STDOUT_SELF,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 3847984c4197SStefano Zampini ierr = MatView(pcbddc->ConstraintMatrix,(PetscViewer)0);CHKERRQ(ierr); 3848984c4197SStefano Zampini */ 3849674ae819SStefano Zampini /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */ 3850674ae819SStefano Zampini if (pcbddc->use_change_of_basis) { 3851026de310SStefano Zampini /* dual and primal dofs on a single cc */ 3852984c4197SStefano Zampini PetscInt dual_dofs,primal_dofs; 3853984c4197SStefano Zampini /* working stuff for GEQRF */ 385481d9aea3SBarry Smith PetscScalar *qr_basis,*qr_tau = NULL,*qr_work,lqr_work_t; 3855984c4197SStefano Zampini PetscBLASInt lqr_work; 3856984c4197SStefano Zampini /* working stuff for UNGQR */ 3857984c4197SStefano Zampini PetscScalar *gqr_work,lgqr_work_t; 3858984c4197SStefano Zampini PetscBLASInt lgqr_work; 3859984c4197SStefano Zampini /* working stuff for TRTRS */ 3860984c4197SStefano Zampini PetscScalar *trs_rhs; 38613f08241aSStefano Zampini PetscBLASInt Blas_NRHS; 3862984c4197SStefano Zampini /* pointers for values insertion into change of basis matrix */ 3863984c4197SStefano Zampini PetscInt *start_rows,*start_cols; 3864984c4197SStefano Zampini PetscScalar *start_vals; 3865984c4197SStefano Zampini /* working stuff for values insertion */ 38664641a718SStefano Zampini PetscBT is_primal; 386764efe560SStefano Zampini PetscInt *aux_primal_numbering_B; 3868906d46d4SStefano Zampini /* matrix sizes */ 3869906d46d4SStefano Zampini PetscInt global_size,local_size; 3870906d46d4SStefano Zampini /* temporary change of basis */ 3871906d46d4SStefano Zampini Mat localChangeOfBasisMatrix; 3872cf5a6209SStefano Zampini /* extra space for debugging */ 3873cf5a6209SStefano Zampini PetscScalar *dbg_work; 3874984c4197SStefano Zampini 3875906d46d4SStefano Zampini /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */ 3876906d46d4SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);CHKERRQ(ierr); 387716f15bc4SStefano Zampini ierr = MatSetType(localChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 3878bbb9e6c6SStefano Zampini ierr = MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);CHKERRQ(ierr); 3879906d46d4SStefano Zampini /* nonzeros for local mat */ 3880bbb9e6c6SStefano Zampini ierr = PetscMalloc1(pcis->n,&nnz);CHKERRQ(ierr); 3881bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) nnz[i]=1; 38829162d606SStefano Zampini for (i=n_vertices;i<total_counts_cc;i++) { 3883a717540cSStefano Zampini if (PetscBTLookup(change_basis,i)) { 38849162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]; 3885a717540cSStefano Zampini if (PetscBTLookup(qr_needed_idx,i)) { 38869162d606SStefano Zampini for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint; 3887a717540cSStefano Zampini } else { 38889162d606SStefano Zampini nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint; 38899162d606SStefano Zampini for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2; 3890a717540cSStefano Zampini } 3891a717540cSStefano Zampini } 3892a717540cSStefano Zampini } 3893906d46d4SStefano Zampini ierr = MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);CHKERRQ(ierr); 3894bbb9e6c6SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 3895a717540cSStefano Zampini /* Set initial identity in the matrix */ 3896bbb9e6c6SStefano Zampini for (i=0;i<pcis->n;i++) { 3897906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);CHKERRQ(ierr); 3898a717540cSStefano Zampini } 3899a717540cSStefano Zampini 3900a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3901a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");CHKERRQ(ierr); 3902a717540cSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 3903a717540cSStefano Zampini } 3904a717540cSStefano Zampini 3905a717540cSStefano Zampini 3906a717540cSStefano Zampini /* Now we loop on the constraints which need a change of basis */ 3907a717540cSStefano Zampini /* 3908a717540cSStefano Zampini Change of basis matrix is evaluated similarly to the FIRST APPROACH in 3909a717540cSStefano Zampini Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1) 3910a717540cSStefano Zampini 3911a6b551f4SStefano Zampini Basic blocks of change of basis matrix T computed by 3912a717540cSStefano Zampini 3913a6b551f4SStefano 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) 3914a6b551f4SStefano Zampini 3915a6b551f4SStefano Zampini | 1 0 ... 0 s_1/S | 3916a6b551f4SStefano Zampini | 0 1 ... 0 s_2/S | 3917a717540cSStefano Zampini | ... | 3918a6b551f4SStefano Zampini | 0 ... 1 s_{n-1}/S | 3919a6b551f4SStefano Zampini | -s_1/s_n ... -s_{n-1}/s_n s_n/S | 3920a717540cSStefano Zampini 3921a6b551f4SStefano Zampini with S = \sum_{i=1}^n s_i^2 3922a6b551f4SStefano Zampini NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering 3923a6b551f4SStefano Zampini in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering 3924a6b551f4SStefano Zampini 3925a6b551f4SStefano Zampini - QR decomposition of constraints otherwise 3926a717540cSStefano Zampini */ 3927a717540cSStefano Zampini if (qr_needed) { 3928984c4197SStefano Zampini /* space to store Q */ 3929854ce69bSBarry Smith ierr = PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);CHKERRQ(ierr); 3930984c4197SStefano Zampini /* first we issue queries for optimal work */ 39313f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39323f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_N);CHKERRQ(ierr); 39333f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3934984c4197SStefano Zampini lqr_work = -1; 39353f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr)); 3936984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr); 3937984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);CHKERRQ(ierr); 3938785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);CHKERRQ(ierr); 3939984c4197SStefano Zampini lgqr_work = -1; 39403f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_M);CHKERRQ(ierr); 39413f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_N);CHKERRQ(ierr); 39423f08241aSStefano Zampini ierr = PetscBLASIntCast(max_constraints,&Blas_K);CHKERRQ(ierr); 39433f08241aSStefano Zampini ierr = PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 39443f08241aSStefano Zampini if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */ 39453f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr)); 3946984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to UNGQR Lapack routine %d",(int)lierr); 3947984c4197SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);CHKERRQ(ierr); 3948785e854fSJed Brown ierr = PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);CHKERRQ(ierr); 3949984c4197SStefano Zampini /* array to store scaling factors for reflectors */ 3950785e854fSJed Brown ierr = PetscMalloc1(max_constraints,&qr_tau);CHKERRQ(ierr); 3951984c4197SStefano Zampini /* array to store rhs and solution of triangular solver */ 3952785e854fSJed Brown ierr = PetscMalloc1(max_constraints*max_constraints,&trs_rhs);CHKERRQ(ierr); 3953a717540cSStefano Zampini /* allocating workspace for check */ 3954a717540cSStefano Zampini if (pcbddc->dbg_flag) { 3955cf5a6209SStefano Zampini ierr = PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);CHKERRQ(ierr); 3956a717540cSStefano Zampini } 3957a717540cSStefano Zampini } 3958984c4197SStefano Zampini /* array to store whether a node is primal or not */ 39594641a718SStefano Zampini ierr = PetscBTCreate(pcis->n_B,&is_primal);CHKERRQ(ierr); 3960473ba861SJed Brown ierr = PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);CHKERRQ(ierr); 39610e6343abSStefano Zampini ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);CHKERRQ(ierr); 396239e2fb2aSStefano Zampini if (i != total_primal_vertices) { 396339e2fb2aSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Error in boundary numbering for BDDC vertices! %d != %d\n",total_primal_vertices,i); 39644641a718SStefano Zampini } 396539e2fb2aSStefano Zampini for (i=0;i<total_primal_vertices;i++) { 396639e2fb2aSStefano Zampini ierr = PetscBTSet(is_primal,aux_primal_numbering_B[i]);CHKERRQ(ierr); 396739e2fb2aSStefano Zampini } 396839e2fb2aSStefano Zampini ierr = PetscFree(aux_primal_numbering_B);CHKERRQ(ierr); 3969984c4197SStefano Zampini 3970a717540cSStefano Zampini /* loop on constraints and see whether or not they need a change of basis and compute it */ 39719162d606SStefano Zampini for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) { 39729162d606SStefano Zampini size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts]; 39734641a718SStefano Zampini if (PetscBTLookup(change_basis,total_counts)) { 3974984c4197SStefano Zampini /* get constraint info */ 39759162d606SStefano Zampini primal_dofs = constraints_n[total_counts]; 3976984c4197SStefano Zampini dual_dofs = size_of_constraint-primal_dofs; 3977984c4197SStefano Zampini 3978984c4197SStefano Zampini if (pcbddc->dbg_flag) { 39799162d606SStefano 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); 3980674ae819SStefano Zampini } 3981984c4197SStefano Zampini 3982fa434743SStefano Zampini if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */ 3983a717540cSStefano Zampini 3984a717540cSStefano Zampini /* copy quadrature constraints for change of basis check */ 3985a717540cSStefano Zampini if (pcbddc->dbg_flag) { 39869162d606SStefano Zampini ierr = PetscMemcpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3987a717540cSStefano Zampini } 3988984c4197SStefano Zampini /* copy temporary constraints into larger work vector (in order to store all columns of Q) */ 39899162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 3990984c4197SStefano Zampini 3991984c4197SStefano Zampini /* compute QR decomposition of constraints */ 39923f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 39933f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 39943f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 3995674ae819SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 39963f08241aSStefano Zampini PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr)); 3997984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr); 3998674ae819SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 3999984c4197SStefano Zampini 4000984c4197SStefano Zampini /* explictly compute R^-T */ 4001984c4197SStefano Zampini ierr = PetscMemzero(trs_rhs,primal_dofs*primal_dofs*sizeof(*trs_rhs));CHKERRQ(ierr); 4002984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0; 40033f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40043f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_NRHS);CHKERRQ(ierr); 40053f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40063f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 4007984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40083f08241aSStefano Zampini PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr)); 4009984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr); 4010984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4011984c4197SStefano Zampini 4012a717540cSStefano Zampini /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */ 40133f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40143f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40153f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40163f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4017984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40183f08241aSStefano Zampini PetscStackCallBLAS("LAPACKungqr",LAPACKungqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr)); 4019984c4197SStefano Zampini if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in UNGQR Lapack routine %d",(int)lierr); 4020984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4021984c4197SStefano Zampini 4022984c4197SStefano Zampini /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints 4023984c4197SStefano Zampini i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below) 4024984c4197SStefano Zampini where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */ 40253f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_M);CHKERRQ(ierr); 40263f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_N);CHKERRQ(ierr); 40273f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_K);CHKERRQ(ierr); 40283f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 40293f08241aSStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDB);CHKERRQ(ierr); 40303f08241aSStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDC);CHKERRQ(ierr); 4031984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 40329162d606SStefano 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)); 4033984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 40349162d606SStefano Zampini ierr = PetscMemcpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs*sizeof(PetscScalar));CHKERRQ(ierr); 4035984c4197SStefano Zampini 4036984c4197SStefano Zampini /* insert values in change of basis matrix respecting global ordering of new primal dofs */ 40379162d606SStefano Zampini start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]]; 4038984c4197SStefano Zampini /* insert cols for primal dofs */ 4039984c4197SStefano Zampini for (j=0;j<primal_dofs;j++) { 4040984c4197SStefano Zampini start_vals = &qr_basis[j*size_of_constraint]; 40419162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4042906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4043984c4197SStefano Zampini } 4044984c4197SStefano Zampini /* insert cols for dual dofs */ 4045984c4197SStefano Zampini for (j=0,k=0;j<dual_dofs;k++) { 40469162d606SStefano Zampini if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) { 4047984c4197SStefano Zampini start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint]; 40489162d606SStefano Zampini start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4049906d46d4SStefano Zampini ierr = MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);CHKERRQ(ierr); 4050984c4197SStefano Zampini j++; 4051674ae819SStefano Zampini } 4052674ae819SStefano Zampini } 4053984c4197SStefano Zampini 4054984c4197SStefano Zampini /* check change of basis */ 4055984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4056984c4197SStefano Zampini PetscInt ii,jj; 4057984c4197SStefano Zampini PetscBool valid_qr=PETSC_TRUE; 4058c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_M);CHKERRQ(ierr); 4059c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 4060c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_K);CHKERRQ(ierr); 4061c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDA);CHKERRQ(ierr); 4062c4303822SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_LDB);CHKERRQ(ierr); 4063c4303822SStefano Zampini ierr = PetscBLASIntCast(primal_dofs,&Blas_LDC);CHKERRQ(ierr); 4064984c4197SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 4065cf5a6209SStefano 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)); 4066984c4197SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 4067984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4068984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4069cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE; 4070cf5a6209SStefano 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; 4071674ae819SStefano Zampini } 4072674ae819SStefano Zampini } 4073984c4197SStefano Zampini if (!valid_qr) { 407422d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");CHKERRQ(ierr); 4075984c4197SStefano Zampini for (jj=0;jj<size_of_constraint;jj++) { 4076984c4197SStefano Zampini for (ii=0;ii<primal_dofs;ii++) { 4077cf5a6209SStefano Zampini if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) { 4078cf5a6209SStefano 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])); 4079674ae819SStefano Zampini } 4080cf5a6209SStefano Zampini if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-1.0) > 1.e-12) { 4081cf5a6209SStefano 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])); 4082984c4197SStefano Zampini } 4083984c4197SStefano Zampini } 4084984c4197SStefano Zampini } 4085674ae819SStefano Zampini } else { 408622d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");CHKERRQ(ierr); 4087674ae819SStefano Zampini } 4088674ae819SStefano Zampini } 4089a717540cSStefano Zampini } else { /* simple transformation block */ 4090a717540cSStefano Zampini PetscInt row,col; 4091a6b551f4SStefano Zampini PetscScalar val,norm; 4092a6b551f4SStefano Zampini 4093a6b551f4SStefano Zampini ierr = PetscBLASIntCast(size_of_constraint,&Blas_N);CHKERRQ(ierr); 40949162d606SStefano 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)); 4095a717540cSStefano Zampini for (j=0;j<size_of_constraint;j++) { 40969162d606SStefano Zampini PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j]; 40979162d606SStefano Zampini row = constraints_idxs[constraints_idxs_ptr[total_counts]+j]; 4098bbb9e6c6SStefano Zampini if (!PetscBTLookup(is_primal,row_B)) { 40999162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]]; 4100906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);CHKERRQ(ierr); 41019162d606SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);CHKERRQ(ierr); 4102a717540cSStefano Zampini } else { 4103a717540cSStefano Zampini for (k=0;k<size_of_constraint;k++) { 41049162d606SStefano Zampini col = constraints_idxs[constraints_idxs_ptr[total_counts]+k]; 4105a717540cSStefano Zampini if (row != col) { 41069162d606SStefano Zampini val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]]; 4107a717540cSStefano Zampini } else { 41089162d606SStefano Zampini val = constraints_data[constraints_data_ptr[total_counts]]/norm; 4109a717540cSStefano Zampini } 4110906d46d4SStefano Zampini ierr = MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);CHKERRQ(ierr); 4111a717540cSStefano Zampini } 4112a717540cSStefano Zampini } 4113a717540cSStefano Zampini } 411498a51de6SStefano Zampini if (pcbddc->dbg_flag) { 411522d5777bSStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");CHKERRQ(ierr); 4116a717540cSStefano Zampini } 4117674ae819SStefano Zampini } 4118984c4197SStefano Zampini } else { 4119984c4197SStefano Zampini if (pcbddc->dbg_flag) { 41209162d606SStefano 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); 4121674ae819SStefano Zampini } 4122674ae819SStefano Zampini } 4123674ae819SStefano Zampini } 4124a717540cSStefano Zampini 4125a717540cSStefano Zampini /* free workspace */ 4126a717540cSStefano Zampini if (qr_needed) { 4127984c4197SStefano Zampini if (pcbddc->dbg_flag) { 4128cf5a6209SStefano Zampini ierr = PetscFree(dbg_work);CHKERRQ(ierr); 4129984c4197SStefano Zampini } 4130984c4197SStefano Zampini ierr = PetscFree(trs_rhs);CHKERRQ(ierr); 4131984c4197SStefano Zampini ierr = PetscFree(qr_tau);CHKERRQ(ierr); 4132984c4197SStefano Zampini ierr = PetscFree(qr_work);CHKERRQ(ierr); 4133984c4197SStefano Zampini ierr = PetscFree(gqr_work);CHKERRQ(ierr); 4134984c4197SStefano Zampini ierr = PetscFree(qr_basis);CHKERRQ(ierr); 4135674ae819SStefano Zampini } 4136a717540cSStefano Zampini ierr = PetscBTDestroy(&is_primal);CHKERRQ(ierr); 4137906d46d4SStefano Zampini ierr = MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4138906d46d4SStefano Zampini ierr = MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4139906d46d4SStefano Zampini 4140906d46d4SStefano Zampini /* assembling of global change of variable */ 414188c03ad3SStefano Zampini if (!pcbddc->fake_change) { 4142bbb9e6c6SStefano Zampini Mat tmat; 414316f15bc4SStefano Zampini PetscInt bs; 414416f15bc4SStefano Zampini 4145906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4146906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4147bbb9e6c6SStefano Zampini ierr = MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);CHKERRQ(ierr); 4148bbb9e6c6SStefano Zampini ierr = MatISSetLocalMat(tmat,localChangeOfBasisMatrix);CHKERRQ(ierr); 4149bbb9e6c6SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4150bbb9e6c6SStefano Zampini ierr = MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);CHKERRQ(ierr); 415116f15bc4SStefano Zampini ierr = MatGetBlockSize(pc->pmat,&bs);CHKERRQ(ierr); 415216f15bc4SStefano Zampini ierr = MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);CHKERRQ(ierr); 4153906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4154bbb9e6c6SStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);CHKERRQ(ierr); 4155bbb9e6c6SStefano Zampini ierr = MatISGetMPIXAIJ(tmat,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4156bbb9e6c6SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4157bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 4158bbb9e6c6SStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 4159e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4160e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4161bbb9e6c6SStefano Zampini ierr = VecReciprocal(pcis->vec1_global);CHKERRQ(ierr); 4162bbb9e6c6SStefano Zampini ierr = MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);CHKERRQ(ierr); 416388c03ad3SStefano Zampini 4164906d46d4SStefano Zampini /* check */ 4165906d46d4SStefano Zampini if (pcbddc->dbg_flag) { 4166906d46d4SStefano Zampini PetscReal error; 4167906d46d4SStefano Zampini Vec x,x_change; 4168906d46d4SStefano Zampini 4169906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x);CHKERRQ(ierr); 4170906d46d4SStefano Zampini ierr = VecDuplicate(pcis->vec1_global,&x_change);CHKERRQ(ierr); 4171906d46d4SStefano Zampini ierr = VecSetRandom(x,NULL);CHKERRQ(ierr); 4172906d46d4SStefano Zampini ierr = VecCopy(x,pcis->vec1_global);CHKERRQ(ierr); 4173e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4174e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 4175bbb9e6c6SStefano Zampini ierr = MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);CHKERRQ(ierr); 4176e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4177e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 4178906d46d4SStefano Zampini ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);CHKERRQ(ierr); 4179906d46d4SStefano Zampini ierr = VecAXPY(x,-1.0,x_change);CHKERRQ(ierr); 4180906d46d4SStefano Zampini ierr = VecNorm(x,NORM_INFINITY,&error);CHKERRQ(ierr); 4181906d46d4SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4182bbb9e6c6SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Error global vs local change: %1.6e\n",error);CHKERRQ(ierr); 4183906d46d4SStefano Zampini ierr = VecDestroy(&x);CHKERRQ(ierr); 4184906d46d4SStefano Zampini ierr = VecDestroy(&x_change);CHKERRQ(ierr); 4185906d46d4SStefano Zampini } 4186b96c3477SStefano Zampini /* adapt sub_schurs computed (if any) */ 4187b96c3477SStefano Zampini if (pcbddc->use_deluxe_scaling) { 4188b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 4189bf3a8328SStefano Zampini 4190bf3a8328SStefano Zampini if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) { 4191bf3a8328SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");CHKERRQ(ierr); 4192bf3a8328SStefano Zampini } 4193b96c3477SStefano Zampini if (sub_schurs->S_Ej_all) { 4194ac632422SStefano Zampini Mat S_new,tmat; 4195bf3a8328SStefano Zampini IS is_all_N,is_V_Sall = NULL; 4196bbb9e6c6SStefano Zampini 4197bbb9e6c6SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);CHKERRQ(ierr); 41986816873aSStefano Zampini ierr = MatGetSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr); 4199bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4200bf3a8328SStefano Zampini ISLocalToGlobalMapping NtoSall; 4201bf3a8328SStefano Zampini IS is_V; 4202b087196eSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);CHKERRQ(ierr); 4203b087196eSStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);CHKERRQ(ierr); 4204b087196eSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);CHKERRQ(ierr); 4205b087196eSStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoSall);CHKERRQ(ierr); 4206b087196eSStefano Zampini ierr = ISDestroy(&is_V);CHKERRQ(ierr); 4207bf3a8328SStefano Zampini } 4208bf3a8328SStefano Zampini ierr = ISDestroy(&is_all_N);CHKERRQ(ierr); 4209ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4210b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 4211ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4212bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4213bf3a8328SStefano Zampini const PetscScalar *array; 4214bf3a8328SStefano Zampini const PetscInt *idxs_V,*idxs_all; 4215bf3a8328SStefano Zampini PetscInt i,n_V; 4216bf3a8328SStefano Zampini 4217b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4218b087196eSStefano Zampini ierr = ISGetLocalSize(is_V_Sall,&n_V);CHKERRQ(ierr); 4219b087196eSStefano Zampini ierr = ISGetIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4220b087196eSStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4221b087196eSStefano Zampini ierr = VecGetArrayRead(pcis->D,&array);CHKERRQ(ierr); 4222b087196eSStefano Zampini for (i=0;i<n_V;i++) { 4223b087196eSStefano Zampini PetscScalar val; 4224b087196eSStefano Zampini PetscInt idx; 4225b087196eSStefano Zampini 4226b087196eSStefano Zampini idx = idxs_V[i]; 4227b087196eSStefano Zampini val = array[idxs_all[idxs_V[i]]]; 4228b087196eSStefano Zampini ierr = MatSetValue(S_new,idx,idx,val,INSERT_VALUES);CHKERRQ(ierr); 4229b087196eSStefano Zampini } 4230b087196eSStefano Zampini ierr = MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4231b087196eSStefano Zampini ierr = MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4232bf3a8328SStefano Zampini ierr = VecRestoreArrayRead(pcis->D,&array);CHKERRQ(ierr); 4233bf3a8328SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);CHKERRQ(ierr); 4234bf3a8328SStefano Zampini ierr = ISRestoreIndices(is_V_Sall,&idxs_V);CHKERRQ(ierr); 4235bf3a8328SStefano Zampini } 4236ac632422SStefano Zampini sub_schurs->S_Ej_all = S_new; 4237ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4238ac632422SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 4239ac632422SStefano Zampini ierr = MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);CHKERRQ(ierr); 4240b96c3477SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 4241ac632422SStefano Zampini ierr = PetscObjectReference((PetscObject)S_new);CHKERRQ(ierr); 4242bf3a8328SStefano Zampini if (pcbddc->deluxe_zerorows) { 4243b087196eSStefano Zampini ierr = MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);CHKERRQ(ierr); 4244bf3a8328SStefano Zampini } 4245ac632422SStefano Zampini sub_schurs->sum_S_Ej_all = S_new; 4246ac632422SStefano Zampini ierr = MatDestroy(&S_new);CHKERRQ(ierr); 4247ac632422SStefano Zampini } 4248b087196eSStefano Zampini ierr = ISDestroy(&is_V_Sall);CHKERRQ(ierr); 424988c03ad3SStefano Zampini ierr = MatDestroy(&tmat);CHKERRQ(ierr); 4250b96c3477SStefano Zampini } 4251b96c3477SStefano Zampini } 4252906d46d4SStefano Zampini ierr = MatDestroy(&localChangeOfBasisMatrix);CHKERRQ(ierr); 425372b8c272SStefano Zampini } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */ 425472b8c272SStefano Zampini ierr = MatDestroy(&pcbddc->ConstraintMatrix);CHKERRQ(ierr); 425572b8c272SStefano Zampini pcbddc->ConstraintMatrix = localChangeOfBasisMatrix; 425672b8c272SStefano Zampini pcbddc->use_qr_single = qr_needed; 425772b8c272SStefano Zampini } 4258906d46d4SStefano Zampini } else if (pcbddc->user_ChangeOfBasisMatrix) { 4259b9b85e73SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4260b9b85e73SStefano Zampini pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix; 4261b9b85e73SStefano Zampini } 4262906d46d4SStefano Zampini 4263906d46d4SStefano Zampini /* set up change of basis context */ 4264906d46d4SStefano Zampini if (pcbddc->ChangeOfBasisMatrix) { 4265906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 4266906d46d4SStefano Zampini 4267906d46d4SStefano Zampini if (!pcbddc->new_global_mat) { 4268906d46d4SStefano Zampini PetscInt global_size,local_size; 4269906d46d4SStefano Zampini 4270906d46d4SStefano Zampini ierr = VecGetSize(pcis->vec1_global,&global_size);CHKERRQ(ierr); 4271906d46d4SStefano Zampini ierr = VecGetLocalSize(pcis->vec1_global,&local_size);CHKERRQ(ierr); 4272906d46d4SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->new_global_mat);CHKERRQ(ierr); 4273906d46d4SStefano Zampini ierr = MatSetSizes(pcbddc->new_global_mat,local_size,local_size,global_size,global_size);CHKERRQ(ierr); 4274906d46d4SStefano Zampini ierr = MatSetType(pcbddc->new_global_mat,MATSHELL);CHKERRQ(ierr); 4275906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT,(void (*)(void))PCBDDCMatMult_Private);CHKERRQ(ierr); 4276906d46d4SStefano Zampini ierr = MatShellSetOperation(pcbddc->new_global_mat,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCMatMultTranspose_Private);CHKERRQ(ierr); 4277906d46d4SStefano Zampini ierr = PetscNew(&change_ctx);CHKERRQ(ierr); 4278906d46d4SStefano Zampini ierr = MatShellSetContext(pcbddc->new_global_mat,change_ctx);CHKERRQ(ierr); 4279906d46d4SStefano Zampini } else { 4280906d46d4SStefano Zampini ierr = MatShellGetContext(pcbddc->new_global_mat,&change_ctx);CHKERRQ(ierr); 4281906d46d4SStefano Zampini ierr = MatDestroy(&change_ctx->global_change);CHKERRQ(ierr); 4282906d46d4SStefano Zampini ierr = VecDestroyVecs(2,&change_ctx->work);CHKERRQ(ierr); 4283906d46d4SStefano Zampini } 4284906d46d4SStefano Zampini if (!pcbddc->user_ChangeOfBasisMatrix) { 4285906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->ChangeOfBasisMatrix);CHKERRQ(ierr); 4286906d46d4SStefano Zampini change_ctx->global_change = pcbddc->ChangeOfBasisMatrix; 4287906d46d4SStefano Zampini } else { 4288906d46d4SStefano Zampini ierr = PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);CHKERRQ(ierr); 4289906d46d4SStefano Zampini change_ctx->global_change = pcbddc->user_ChangeOfBasisMatrix; 4290906d46d4SStefano Zampini } 4291906d46d4SStefano Zampini ierr = VecDuplicateVecs(pcis->vec1_global,2,&change_ctx->work);CHKERRQ(ierr); 4292906d46d4SStefano Zampini ierr = MatSetUp(pcbddc->new_global_mat);CHKERRQ(ierr); 4293906d46d4SStefano Zampini ierr = MatAssemblyBegin(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4294906d46d4SStefano Zampini ierr = MatAssemblyEnd(pcbddc->new_global_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4295a8661815SStefano Zampini } else { 4296a8661815SStefano Zampini ierr = MatDestroy(&pcbddc->new_global_mat);CHKERRQ(ierr); 4297b9b85e73SStefano Zampini } 4298a717540cSStefano Zampini 429972b8c272SStefano Zampini if (!pcbddc->fake_change) { 43004f1b2e48SStefano Zampini /* add pressure dofs to set of primal nodes for numbering purposes */ 43014f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 43024f1b2e48SStefano Zampini pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i]; 43034f1b2e48SStefano Zampini pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i]; 4304019a44ceSStefano Zampini pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1; 4305019a44ceSStefano Zampini pcbddc->local_primal_size_cc++; 4306019a44ceSStefano Zampini pcbddc->local_primal_size++; 4307019a44ceSStefano Zampini } 4308019a44ceSStefano Zampini 4309019a44ceSStefano Zampini /* check if a new primal space has been introduced (also take into account benign trick) */ 4310727cdba6SStefano Zampini pcbddc->new_primal_space_local = PETSC_TRUE; 4311727cdba6SStefano Zampini if (olocal_primal_size == pcbddc->local_primal_size) { 43129f47a83aSStefano 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); 4313c1c8e736SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 43140e6343abSStefano Zampini if (!pcbddc->new_primal_space_local) { 43159f47a83aSStefano 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); 4316727cdba6SStefano Zampini pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local); 4317727cdba6SStefano Zampini } 43180e6343abSStefano Zampini } 4319727cdba6SStefano Zampini /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */ 4320727cdba6SStefano Zampini ierr = MPI_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 432172b8c272SStefano Zampini } 432272b8c272SStefano Zampini ierr = PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);CHKERRQ(ierr); 4323727cdba6SStefano Zampini 4324a717540cSStefano Zampini /* flush dbg viewer */ 4325b8ffe317SStefano Zampini if (pcbddc->dbg_flag) { 4326b8ffe317SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4327b8ffe317SStefano Zampini } 4328a717540cSStefano Zampini 4329e310c8b4SStefano Zampini /* free workspace */ 4330a717540cSStefano Zampini ierr = PetscBTDestroy(&qr_needed_idx);CHKERRQ(ierr); 43314641a718SStefano Zampini ierr = PetscBTDestroy(&change_basis);CHKERRQ(ierr); 433208122e43SStefano Zampini if (!pcbddc->adaptive_selection) { 43339162d606SStefano Zampini ierr = PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);CHKERRQ(ierr); 43349162d606SStefano Zampini ierr = PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);CHKERRQ(ierr); 433508122e43SStefano Zampini } else { 43369162d606SStefano Zampini ierr = PetscFree5(pcbddc->adaptive_constraints_n, 43379162d606SStefano Zampini pcbddc->adaptive_constraints_idxs_ptr, 43389162d606SStefano Zampini pcbddc->adaptive_constraints_data_ptr, 433908122e43SStefano Zampini pcbddc->adaptive_constraints_idxs, 434008122e43SStefano Zampini pcbddc->adaptive_constraints_data);CHKERRQ(ierr); 43419162d606SStefano Zampini ierr = PetscFree(constraints_n);CHKERRQ(ierr); 43429162d606SStefano Zampini ierr = PetscFree(constraints_idxs_B);CHKERRQ(ierr); 434308122e43SStefano Zampini } 4344674ae819SStefano Zampini PetscFunctionReturn(0); 4345674ae819SStefano Zampini } 4346674ae819SStefano Zampini 4347674ae819SStefano Zampini #undef __FUNCT__ 4348674ae819SStefano Zampini #define __FUNCT__ "PCBDDCAnalyzeInterface" 4349674ae819SStefano Zampini PetscErrorCode PCBDDCAnalyzeInterface(PC pc) 4350674ae819SStefano Zampini { 4351674ae819SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 4352674ae819SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 4353674ae819SStefano Zampini Mat_IS *matis = (Mat_IS*)pc->pmat->data; 435414f95afaSStefano Zampini PetscInt ierr,i,N; 4355674ae819SStefano Zampini PetscViewer viewer=pcbddc->dbg_viewer; 4356674ae819SStefano Zampini 4357674ae819SStefano Zampini PetscFunctionBegin; 43588e61c736SStefano Zampini /* Reset previously computed graph */ 43598e61c736SStefano Zampini ierr = PCBDDCGraphReset(pcbddc->mat_graph);CHKERRQ(ierr); 4360674ae819SStefano Zampini /* Init local Graph struct */ 43617fb0e2dbSStefano Zampini ierr = MatGetSize(pc->pmat,&N,NULL);CHKERRQ(ierr); 43623bbff08aSStefano Zampini ierr = PCBDDCGraphInit(pcbddc->mat_graph,pcis->mapping,N);CHKERRQ(ierr); 4363674ae819SStefano Zampini 4364575ad6abSStefano Zampini /* Check validity of the csr graph passed in by the user */ 43655099eff2SStefano Zampini if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) { 43665099eff2SStefano 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); 4367575ad6abSStefano Zampini } 43689577ea80SStefano Zampini 4369674ae819SStefano Zampini /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */ 4370d4d8cf7bSStefano Zampini if ( (!pcbddc->mat_graph->xadj || !pcbddc->mat_graph->adjncy) && pcbddc->use_local_adj) { 43714d379d7bSStefano Zampini PetscInt *xadj,*adjncy; 43724d379d7bSStefano Zampini PetscInt nvtxs; 4373e496cd5dSStefano Zampini PetscBool flg_row=PETSC_FALSE; 4374674ae819SStefano Zampini 43752fffb893SStefano Zampini ierr = MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 43762fffb893SStefano Zampini if (flg_row) { 43774d379d7bSStefano Zampini ierr = PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);CHKERRQ(ierr); 4378b96c3477SStefano Zampini pcbddc->computed_rowadj = PETSC_TRUE; 43792fffb893SStefano Zampini } 43802fffb893SStefano Zampini ierr = MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);CHKERRQ(ierr); 4381674ae819SStefano Zampini } 43829b28b941SStefano Zampini if (pcbddc->dbg_flag) { 43839b28b941SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 4384674ae819SStefano Zampini } 4385674ae819SStefano Zampini 438663602bcaSStefano Zampini /* Set default dofs' splitting if no information has been provided by the user with PCBDDCSetDofsSplitting or PCBDDCSetDofsSplittingLocal */ 438763602bcaSStefano Zampini if (pcbddc->user_provided_isfordofs) { 438863602bcaSStefano Zampini if (pcbddc->n_ISForDofs) { /* need to convert from global to local and remove references to global dofs splitting */ 438995ecbf38SStefano Zampini ierr = PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 439063602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofs;i++) { 4391e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 439263602bcaSStefano Zampini ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr); 4393674ae819SStefano Zampini } 439463602bcaSStefano Zampini pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs; 439563602bcaSStefano Zampini pcbddc->n_ISForDofs = 0; 439663602bcaSStefano Zampini ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr); 4397674ae819SStefano Zampini } 439863602bcaSStefano Zampini } else { 439963602bcaSStefano Zampini if (!pcbddc->n_ISForDofsLocal) { /* field split not present, create it in local ordering */ 440063602bcaSStefano Zampini ierr = MatGetBlockSize(pc->pmat,&pcbddc->n_ISForDofsLocal);CHKERRQ(ierr); 4401854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);CHKERRQ(ierr); 440263602bcaSStefano Zampini for (i=0;i<pcbddc->n_ISForDofsLocal;i++) { 440363602bcaSStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)pc),pcis->n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);CHKERRQ(ierr); 440463602bcaSStefano Zampini } 440563602bcaSStefano Zampini } 4406674ae819SStefano Zampini } 4407674ae819SStefano Zampini 4408674ae819SStefano Zampini /* Setup of Graph */ 4409785d1243SStefano Zampini if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) { /* need to convert from global to local */ 4410e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);CHKERRQ(ierr); 4411785d1243SStefano Zampini } 4412785d1243SStefano Zampini if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) { /* need to convert from global to local */ 4413e176bc59SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);CHKERRQ(ierr); 4414785d1243SStefano Zampini } 441530368db7SStefano Zampini if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) { /* need to convert from global to local */ 441630368db7SStefano Zampini ierr = PCBDDCGlobalToLocal(matis->rctx,pcis->vec1_global,pcis->vec1_N,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);CHKERRQ(ierr); 441730368db7SStefano Zampini } 441814f95afaSStefano 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); 4419674ae819SStefano Zampini 44204f1b2e48SStefano Zampini /* attach info on disconnected subdomains if present */ 44214f1b2e48SStefano Zampini if (pcbddc->n_local_subs) { 44224f1b2e48SStefano Zampini PetscInt *local_subs; 44234f1b2e48SStefano Zampini 44244f1b2e48SStefano Zampini ierr = PetscMalloc1(N,&local_subs);CHKERRQ(ierr); 44254f1b2e48SStefano Zampini for (i=0;i<pcbddc->n_local_subs;i++) { 44264f1b2e48SStefano Zampini const PetscInt *idxs; 44274f1b2e48SStefano Zampini PetscInt nl,j; 44284f1b2e48SStefano Zampini 44294f1b2e48SStefano Zampini ierr = ISGetLocalSize(pcbddc->local_subs[i],&nl);CHKERRQ(ierr); 44304f1b2e48SStefano Zampini ierr = ISGetIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44314f1b2e48SStefano Zampini for (j=0;j<nl;j++) { 44324f1b2e48SStefano Zampini local_subs[idxs[j]] = i; 44334f1b2e48SStefano Zampini } 44344f1b2e48SStefano Zampini ierr = ISRestoreIndices(pcbddc->local_subs[i],&idxs);CHKERRQ(ierr); 44354f1b2e48SStefano Zampini } 44364f1b2e48SStefano Zampini pcbddc->mat_graph->n_local_subs = pcbddc->n_local_subs; 44374f1b2e48SStefano Zampini pcbddc->mat_graph->local_subs = local_subs; 44384f1b2e48SStefano Zampini } 44394f1b2e48SStefano Zampini 4440674ae819SStefano Zampini /* Graph's connected components analysis */ 4441674ae819SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);CHKERRQ(ierr); 4442674ae819SStefano Zampini 4443674ae819SStefano Zampini /* print some info to stdout */ 4444674ae819SStefano Zampini if (pcbddc->dbg_flag) { 4445302440fdSBarry Smith ierr = PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,viewer);CHKERRQ(ierr); 4446674ae819SStefano Zampini } 4447674ae819SStefano Zampini PetscFunctionReturn(0); 4448674ae819SStefano Zampini } 4449674ae819SStefano Zampini 4450dc456d91SStefano Zampini /* given an index sets possibly with holes, renumbers the indexes removing the holes */ 4451674ae819SStefano Zampini #undef __FUNCT__ 4452674ae819SStefano Zampini #define __FUNCT__ "PCBDDCSubsetNumbering" 4453dc456d91SStefano Zampini PetscErrorCode PCBDDCSubsetNumbering(IS subset, IS subset_mult, PetscInt *N_n, IS *subset_n) 4454674ae819SStefano Zampini { 4455dc456d91SStefano Zampini PetscSF sf; 4456dc456d91SStefano Zampini PetscLayout map; 4457dc456d91SStefano Zampini const PetscInt *idxs; 4458dc456d91SStefano Zampini PetscInt *leaf_data,*root_data,*gidxs; 4459dc456d91SStefano Zampini PetscInt N,n,i,lbounds[2],gbounds[2],Nl; 4460dc456d91SStefano Zampini PetscInt n_n,nlocals,start,first_index; 4461dc456d91SStefano Zampini PetscMPIInt commsize; 4462674ae819SStefano Zampini PetscBool first_found; 4463674ae819SStefano Zampini PetscErrorCode ierr; 4464674ae819SStefano Zampini 4465674ae819SStefano Zampini PetscFunctionBegin; 4466dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&n);CHKERRQ(ierr); 4467dc456d91SStefano Zampini if (subset_mult) { 4468dc456d91SStefano Zampini PetscCheckSameComm(subset,1,subset_mult,2); 4469dc456d91SStefano Zampini ierr = ISGetLocalSize(subset,&i);CHKERRQ(ierr); 4470dc456d91SStefano Zampini if (i != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Local subset and multiplicity sizes don't match! %d != %d",n,i); 4471674ae819SStefano Zampini } 4472dc456d91SStefano Zampini /* create workspace layout for computing global indices of subset */ 4473dc456d91SStefano Zampini ierr = ISGetIndices(subset,&idxs);CHKERRQ(ierr); 4474dc456d91SStefano Zampini lbounds[0] = lbounds[1] = 0; 4475dc456d91SStefano Zampini for (i=0;i<n;i++) { 4476dc456d91SStefano Zampini if (idxs[i] < lbounds[0]) lbounds[0] = idxs[i]; 4477dc456d91SStefano Zampini else if (idxs[i] > lbounds[1]) lbounds[1] = idxs[i]; 4478674ae819SStefano Zampini } 4479dc456d91SStefano Zampini lbounds[0] = -lbounds[0]; 4480dc456d91SStefano Zampini ierr = MPI_Allreduce(lbounds,gbounds,2,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4481dc456d91SStefano Zampini gbounds[0] = -gbounds[0]; 4482dc456d91SStefano Zampini N = gbounds[1] - gbounds[0] + 1; 4483dc456d91SStefano Zampini ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)subset),&map);CHKERRQ(ierr); 4484dc456d91SStefano Zampini ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr); 4485dc456d91SStefano Zampini ierr = PetscLayoutSetSize(map,N);CHKERRQ(ierr); 4486dc456d91SStefano Zampini ierr = PetscLayoutSetUp(map);CHKERRQ(ierr); 4487dc456d91SStefano Zampini ierr = PetscLayoutGetLocalSize(map,&Nl);CHKERRQ(ierr); 4488dc456d91SStefano Zampini 4489dc456d91SStefano Zampini /* create sf : leaf_data == multiplicity of indexes, root data == global index in layout */ 4490dc456d91SStefano Zampini ierr = PetscMalloc2(n,&leaf_data,Nl,&root_data);CHKERRQ(ierr); 4491dc456d91SStefano Zampini if (subset_mult) { 4492dc456d91SStefano Zampini const PetscInt* idxs_mult; 4493dc456d91SStefano Zampini 4494dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4495dc456d91SStefano Zampini ierr = PetscMemcpy(leaf_data,idxs_mult,n*sizeof(PetscInt));CHKERRQ(ierr); 4496dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4497674ae819SStefano Zampini } else { 4498dc456d91SStefano Zampini for (i=0;i<n;i++) leaf_data[i] = 1; 4499674ae819SStefano Zampini } 4500dc456d91SStefano Zampini /* local size of new subset */ 4501dc456d91SStefano Zampini n_n = 0; 4502dc456d91SStefano Zampini for (i=0;i<n;i++) n_n += leaf_data[i]; 4503dc456d91SStefano Zampini 4504dc456d91SStefano Zampini /* global indexes in layout */ 4505dc456d91SStefano Zampini ierr = PetscMalloc1(n_n,&gidxs);CHKERRQ(ierr); /* allocating possibly extra space in gidxs which will be used later */ 4506dc456d91SStefano Zampini for (i=0;i<n;i++) gidxs[i] = idxs[i] - gbounds[0]; 4507dc456d91SStefano Zampini ierr = ISRestoreIndices(subset,&idxs);CHKERRQ(ierr); 4508dc456d91SStefano Zampini ierr = PetscSFCreate(PetscObjectComm((PetscObject)subset),&sf);CHKERRQ(ierr); 4509dc456d91SStefano Zampini ierr = PetscSFSetGraphLayout(sf,map,n,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr); 4510dc456d91SStefano Zampini ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr); 4511dc456d91SStefano Zampini 4512dc456d91SStefano Zampini /* reduce from leaves to roots */ 4513dc456d91SStefano Zampini ierr = PetscMemzero(root_data,Nl*sizeof(PetscInt));CHKERRQ(ierr); 451464a8e5bcSStefano Zampini ierr = PetscSFReduceBegin(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 451564a8e5bcSStefano Zampini ierr = PetscSFReduceEnd(sf,MPIU_INT,leaf_data,root_data,MPI_MAX);CHKERRQ(ierr); 4516dc456d91SStefano Zampini 4517dc456d91SStefano Zampini /* count indexes in local part of layout */ 4518674ae819SStefano Zampini nlocals = 0; 4519674ae819SStefano Zampini first_index = -1; 4520674ae819SStefano Zampini first_found = PETSC_FALSE; 4521dc456d91SStefano Zampini for (i=0;i<Nl;i++) { 4522dc456d91SStefano Zampini if (!first_found && root_data[i]) { 4523674ae819SStefano Zampini first_found = PETSC_TRUE; 4524674ae819SStefano Zampini first_index = i; 4525674ae819SStefano Zampini } 4526dc456d91SStefano Zampini nlocals += root_data[i]; 4527674ae819SStefano Zampini } 4528dc456d91SStefano Zampini 4529dc456d91SStefano Zampini /* cumulative of number of indexes and size of subset without holes */ 45305fa240b1SStefano Zampini #if defined(PETSC_HAVE_MPI_EXSCAN) 4531dc456d91SStefano Zampini start = 0; 453264a8e5bcSStefano Zampini ierr = MPI_Exscan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45335fa240b1SStefano Zampini #else 453464a8e5bcSStefano Zampini ierr = MPI_Scan(&nlocals,&start,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 45355fa240b1SStefano Zampini start = start-nlocals; 45365fa240b1SStefano Zampini #endif 45375fa240b1SStefano Zampini 4538dc456d91SStefano Zampini if (N_n) { /* compute total size of new subset if requested */ 4539dc456d91SStefano Zampini *N_n = start + nlocals; 4540dc456d91SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)subset),&commsize);CHKERRQ(ierr); 4541dc456d91SStefano Zampini ierr = MPI_Bcast(N_n,1,MPIU_INT,commsize-1,PetscObjectComm((PetscObject)subset));CHKERRQ(ierr); 4542674ae819SStefano Zampini } 45435fa240b1SStefano Zampini 45445fa240b1SStefano Zampini /* adapt root data with cumulative */ 4545674ae819SStefano Zampini if (first_found) { 4546dc456d91SStefano Zampini PetscInt old_index; 4547dc456d91SStefano Zampini 4548dc456d91SStefano Zampini root_data[first_index] += start; 4549674ae819SStefano Zampini old_index = first_index; 4550dc456d91SStefano Zampini for (i=first_index+1;i<Nl;i++) { 4551dc456d91SStefano Zampini if (root_data[i]) { 4552dc456d91SStefano Zampini root_data[i] += root_data[old_index]; 4553674ae819SStefano Zampini old_index = i; 4554674ae819SStefano Zampini } 4555674ae819SStefano Zampini } 4556674ae819SStefano Zampini } 4557dc456d91SStefano Zampini 4558dc456d91SStefano Zampini /* from roots to leaves */ 4559dc456d91SStefano Zampini ierr = PetscSFBcastBegin(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4560dc456d91SStefano Zampini ierr = PetscSFBcastEnd(sf,MPIU_INT,root_data,leaf_data);CHKERRQ(ierr); 4561dc456d91SStefano Zampini ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 4562dc456d91SStefano Zampini 4563dc456d91SStefano Zampini /* create new IS with global indexes without holes */ 4564dc456d91SStefano Zampini if (subset_mult) { 4565dc456d91SStefano Zampini const PetscInt* idxs_mult; 4566dc456d91SStefano Zampini PetscInt cum; 4567dc456d91SStefano Zampini 4568dc456d91SStefano Zampini cum = 0; 4569dc456d91SStefano Zampini ierr = ISGetIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4570dc456d91SStefano Zampini for (i=0;i<n;i++) { 4571dc456d91SStefano Zampini PetscInt j; 4572dc456d91SStefano Zampini for (j=0;j<idxs_mult[i];j++) gidxs[cum++] = leaf_data[i] - idxs_mult[i] + j; 4573674ae819SStefano Zampini } 4574dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_mult,&idxs_mult);CHKERRQ(ierr); 4575674ae819SStefano Zampini } else { 4576dc456d91SStefano Zampini for (i=0;i<n;i++) { 4577dc456d91SStefano Zampini gidxs[i] = leaf_data[i]-1; 4578674ae819SStefano Zampini } 4579674ae819SStefano Zampini } 4580dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)subset),n_n,gidxs,PETSC_OWN_POINTER,subset_n);CHKERRQ(ierr); 4581dc456d91SStefano Zampini ierr = PetscFree2(leaf_data,root_data);CHKERRQ(ierr); 4582674ae819SStefano Zampini PetscFunctionReturn(0); 4583674ae819SStefano Zampini } 45849a7d3425SStefano Zampini 45859a7d3425SStefano Zampini #undef __FUNCT__ 45869a7d3425SStefano Zampini #define __FUNCT__ "PCBDDCOrthonormalizeVecs" 45879a7d3425SStefano Zampini PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt n, Vec vecs[]) 45889a7d3425SStefano Zampini { 45899a7d3425SStefano Zampini PetscInt i,j; 45909a7d3425SStefano Zampini PetscScalar *alphas; 45919a7d3425SStefano Zampini PetscErrorCode ierr; 45929a7d3425SStefano Zampini 45939a7d3425SStefano Zampini PetscFunctionBegin; 45949a7d3425SStefano Zampini /* this implements stabilized Gram-Schmidt */ 4595785e854fSJed Brown ierr = PetscMalloc1(n,&alphas);CHKERRQ(ierr); 45969a7d3425SStefano Zampini for (i=0;i<n;i++) { 45979a7d3425SStefano Zampini ierr = VecNormalize(vecs[i],NULL);CHKERRQ(ierr); 45989a7d3425SStefano Zampini if (i<n) { ierr = VecMDot(vecs[i],n-i-1,&vecs[i+1],&alphas[i+1]);CHKERRQ(ierr); } 45999a7d3425SStefano Zampini for (j=i+1;j<n;j++) { ierr = VecAXPY(vecs[j],PetscConj(-alphas[j]),vecs[i]);CHKERRQ(ierr); } 46009a7d3425SStefano Zampini } 46019a7d3425SStefano Zampini ierr = PetscFree(alphas);CHKERRQ(ierr); 46029a7d3425SStefano Zampini PetscFunctionReturn(0); 46039a7d3425SStefano Zampini } 46049a7d3425SStefano Zampini 4605e7931f94SStefano Zampini #undef __FUNCT__ 460670cf5478SStefano Zampini #define __FUNCT__ "MatISGetSubassemblingPattern" 4607b0c7d250SStefano Zampini PetscErrorCode MatISGetSubassemblingPattern(Mat mat, PetscInt n_subdomains, PetscInt redprocs, IS* is_sends) 4608e7931f94SStefano Zampini { 460952e5ac9dSStefano Zampini IS ranks_send_to; 4610e7931f94SStefano Zampini PetscInt n_neighs,*neighs,*n_shared,**shared; 4611e7931f94SStefano Zampini PetscMPIInt size,rank,color; 461252e5ac9dSStefano Zampini PetscInt *xadj,*adjncy; 461352e5ac9dSStefano Zampini PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx; 46143837a79fSStefano Zampini PetscInt i,local_size,threshold=0; 46152b510759SStefano Zampini PetscBool use_vwgt=PETSC_FALSE,use_square=PETSC_FALSE; 4616e7931f94SStefano Zampini PetscSubcomm subcomm; 461752e5ac9dSStefano Zampini PetscErrorCode ierr; 4618a57a6d2fSStefano Zampini 4619e7931f94SStefano Zampini PetscFunctionBegin; 46202b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_square",&use_square,NULL);CHKERRQ(ierr); 46212b510759SStefano Zampini ierr = PetscOptionsGetBool(NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);CHKERRQ(ierr); 46222b510759SStefano Zampini ierr = PetscOptionsGetInt(NULL,"-matis_partitioning_threshold",&threshold,NULL);CHKERRQ(ierr); 4623e7931f94SStefano Zampini 4624e7931f94SStefano Zampini /* Get info on mapping */ 46253bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&local_size);CHKERRQ(ierr); 46263bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4627e7931f94SStefano Zampini 4628e7931f94SStefano Zampini /* build local CSR graph of subdomains' connectivity */ 4629785e854fSJed Brown ierr = PetscMalloc1(2,&xadj);CHKERRQ(ierr); 4630e7931f94SStefano Zampini xadj[0] = 0; 4631e7931f94SStefano Zampini xadj[1] = PetscMax(n_neighs-1,0); 4632785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy);CHKERRQ(ierr); 4633785e854fSJed Brown ierr = PetscMalloc1(xadj[1],&adjncy_wgt);CHKERRQ(ierr); 4634e7931f94SStefano Zampini 46352b510759SStefano Zampini if (threshold) { 4636d023bfaeSStefano Zampini PetscInt xadj_count = 0; 46372b510759SStefano Zampini for (i=1;i<n_neighs;i++) { 4638d023bfaeSStefano Zampini if (n_shared[i] > threshold) { 4639d023bfaeSStefano Zampini adjncy[xadj_count] = neighs[i]; 4640d023bfaeSStefano Zampini adjncy_wgt[xadj_count] = n_shared[i]; 4641d023bfaeSStefano Zampini xadj_count++; 4642e7931f94SStefano Zampini } 4643e7931f94SStefano Zampini } 4644d023bfaeSStefano Zampini xadj[1] = xadj_count; 4645c8587f34SStefano Zampini } else { 4646e7931f94SStefano Zampini if (xadj[1]) { 4647e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy,&neighs[1],xadj[1]*sizeof(*adjncy));CHKERRQ(ierr); 4648e7931f94SStefano Zampini ierr = PetscMemcpy(adjncy_wgt,&n_shared[1],xadj[1]*sizeof(*adjncy_wgt));CHKERRQ(ierr); 4649c8587f34SStefano Zampini } 4650e7931f94SStefano Zampini } 46513bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);CHKERRQ(ierr); 4652e7931f94SStefano Zampini if (use_square) { 4653e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4654e7931f94SStefano Zampini adjncy_wgt[i] = adjncy_wgt[i]*adjncy_wgt[i]; 4655e7931f94SStefano Zampini } 4656e7931f94SStefano Zampini } 4657e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4658e7931f94SStefano Zampini 46593837a79fSStefano Zampini ierr = PetscMalloc1(1,&ranks_send_to_idx);CHKERRQ(ierr); 4660e7931f94SStefano Zampini 4661e7931f94SStefano Zampini /* 4662e7931f94SStefano Zampini Restrict work on active processes only. 4663e7931f94SStefano Zampini */ 4664e7931f94SStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&subcomm);CHKERRQ(ierr); 4665e7931f94SStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); /* 2 groups, active process and not active processes */ 4666e7931f94SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 46672b510759SStefano Zampini ierr = PetscMPIIntCast(!local_size,&color);CHKERRQ(ierr); 4668d3531aaaSJed Brown ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4669e7931f94SStefano Zampini if (color) { 4670e7931f94SStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 4671e7931f94SStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 4672e7931f94SStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4673c8587f34SStefano Zampini } else { 467452e5ac9dSStefano Zampini Mat subdomain_adj; 467552e5ac9dSStefano Zampini IS new_ranks,new_ranks_contig; 467652e5ac9dSStefano Zampini MatPartitioning partitioner; 467752e5ac9dSStefano Zampini PetscInt prank,rstart=0,rend=0; 467852e5ac9dSStefano Zampini PetscInt *is_indices,*oldranks; 4679b0c7d250SStefano Zampini PetscBool aggregate; 4680b0c7d250SStefano Zampini 4681306c2d5bSBarry Smith ierr = MPI_Comm_size(PetscSubcommChild(subcomm),&size);CHKERRQ(ierr); 4682785e854fSJed Brown ierr = PetscMalloc1(size,&oldranks);CHKERRQ(ierr); 4683e7931f94SStefano Zampini prank = rank; 4684306c2d5bSBarry Smith ierr = MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,PetscSubcommChild(subcomm));CHKERRQ(ierr); 46858002ef2cSStefano Zampini /* 4686e7931f94SStefano Zampini for (i=0;i<size;i++) { 4687e7931f94SStefano Zampini PetscPrintf(subcomm->comm,"oldranks[%d] = %d\n",i,oldranks[i]); 4688c8587f34SStefano Zampini } 46898002ef2cSStefano Zampini */ 4690e7931f94SStefano Zampini for (i=0;i<xadj[1];i++) { 4691e7931f94SStefano Zampini ierr = PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);CHKERRQ(ierr); 4692c8587f34SStefano Zampini } 4693e7931f94SStefano Zampini ierr = PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);CHKERRQ(ierr); 4694b0c7d250SStefano Zampini aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE); 4695b0c7d250SStefano Zampini if (aggregate) { 4696b0c7d250SStefano Zampini PetscInt lrows,row,ncols,*cols; 4697b0c7d250SStefano Zampini PetscMPIInt nrank; 4698b0c7d250SStefano Zampini PetscScalar *vals; 4699b0c7d250SStefano Zampini 4700b0c7d250SStefano Zampini ierr = MPI_Comm_rank(PetscSubcommChild(subcomm),&nrank);CHKERRQ(ierr); 4701b0c7d250SStefano Zampini lrows = 0; 4702b0c7d250SStefano Zampini if (nrank<redprocs) { 4703b0c7d250SStefano Zampini lrows = size/redprocs; 4704b0c7d250SStefano Zampini if (nrank<size%redprocs) lrows++; 4705b0c7d250SStefano Zampini } 47065fa240b1SStefano Zampini ierr = MatCreateAIJ(PetscSubcommChild(subcomm),lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);CHKERRQ(ierr); 4707b0c7d250SStefano Zampini ierr = MatGetOwnershipRange(subdomain_adj,&rstart,&rend);CHKERRQ(ierr); 4708b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4709b0c7d250SStefano Zampini ierr = MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4710b0c7d250SStefano Zampini row = nrank; 4711b0c7d250SStefano Zampini ncols = xadj[1]-xadj[0]; 4712b0c7d250SStefano Zampini cols = adjncy; 4713b0c7d250SStefano Zampini ierr = PetscMalloc1(ncols,&vals);CHKERRQ(ierr); 4714b0c7d250SStefano Zampini for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i]; 4715b0c7d250SStefano Zampini ierr = MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 4716b0c7d250SStefano Zampini ierr = MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4717b0c7d250SStefano Zampini ierr = MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 471852e5ac9dSStefano Zampini ierr = PetscFree(xadj);CHKERRQ(ierr); 471952e5ac9dSStefano Zampini ierr = PetscFree(adjncy);CHKERRQ(ierr); 472052e5ac9dSStefano Zampini ierr = PetscFree(adjncy_wgt);CHKERRQ(ierr); 4721b0c7d250SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 4722b0c7d250SStefano Zampini } else { 4723306c2d5bSBarry Smith ierr = MatCreateMPIAdj(PetscSubcommChild(subcomm),1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);CHKERRQ(ierr); 4724b0c7d250SStefano Zampini } 472522b6e8a2SStefano Zampini /* ierr = MatView(subdomain_adj,0);CHKERRQ(ierr); */ 4726e7931f94SStefano Zampini 4727e7931f94SStefano Zampini /* Partition */ 4728306c2d5bSBarry Smith ierr = MatPartitioningCreate(PetscSubcommChild(subcomm),&partitioner);CHKERRQ(ierr); 4729e7931f94SStefano Zampini ierr = MatPartitioningSetAdjacency(partitioner,subdomain_adj);CHKERRQ(ierr); 4730e7931f94SStefano Zampini if (use_vwgt) { 47313837a79fSStefano Zampini ierr = PetscMalloc1(1,&v_wgt);CHKERRQ(ierr); 4732e7931f94SStefano Zampini v_wgt[0] = local_size; 4733e7931f94SStefano Zampini ierr = MatPartitioningSetVertexWeights(partitioner,v_wgt);CHKERRQ(ierr); 4734c8587f34SStefano Zampini } 473528143c3dSStefano Zampini n_subdomains = PetscMin((PetscInt)size,n_subdomains); 473628143c3dSStefano Zampini ierr = MatPartitioningSetNParts(partitioner,n_subdomains);CHKERRQ(ierr); 4737e7931f94SStefano Zampini ierr = MatPartitioningSetFromOptions(partitioner);CHKERRQ(ierr); 4738e7931f94SStefano Zampini ierr = MatPartitioningApply(partitioner,&new_ranks);CHKERRQ(ierr); 473922b6e8a2SStefano Zampini /* ierr = MatPartitioningView(partitioner,0);CHKERRQ(ierr); */ 4740e7931f94SStefano Zampini 474152e5ac9dSStefano Zampini /* renumber new_ranks to avoid "holes" in new set of processors */ 474252e5ac9dSStefano Zampini ierr = PCBDDCSubsetNumbering(new_ranks,NULL,NULL,&new_ranks_contig);CHKERRQ(ierr); 474352e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks);CHKERRQ(ierr); 474452e5ac9dSStefano Zampini ierr = ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4745b0c7d250SStefano Zampini if (!redprocs) { 4746b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[is_indices[0]]; 474728143c3dSStefano Zampini } else { 4748b0c7d250SStefano Zampini PetscInt idxs[1]; 4749b0c7d250SStefano Zampini PetscMPIInt tag; 4750b0c7d250SStefano Zampini MPI_Request *reqs; 4751b0c7d250SStefano Zampini 4752b0c7d250SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);CHKERRQ(ierr); 4753b0c7d250SStefano Zampini ierr = PetscMalloc1(rend-rstart,&reqs);CHKERRQ(ierr); 4754b0c7d250SStefano Zampini for (i=rstart;i<rend;i++) { 4755b0c7d250SStefano Zampini ierr = MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,PetscSubcommChild(subcomm),&reqs[i-rstart]);CHKERRQ(ierr); 475628143c3dSStefano Zampini } 4757b0c7d250SStefano Zampini ierr = MPI_Recv(idxs,1,MPIU_INT,MPI_ANY_SOURCE,tag,PetscSubcommChild(subcomm),MPI_STATUS_IGNORE);CHKERRQ(ierr); 4758b0c7d250SStefano Zampini ierr = MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 4759b0c7d250SStefano Zampini ierr = PetscFree(reqs);CHKERRQ(ierr); 4760b0c7d250SStefano Zampini ranks_send_to_idx[0] = oldranks[idxs[0]]; 4761e7931f94SStefano Zampini } 476252e5ac9dSStefano Zampini ierr = ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);CHKERRQ(ierr); 4763e7931f94SStefano Zampini /* clean up */ 4764e7931f94SStefano Zampini ierr = PetscFree(oldranks);CHKERRQ(ierr); 476552e5ac9dSStefano Zampini ierr = ISDestroy(&new_ranks_contig);CHKERRQ(ierr); 4766e7931f94SStefano Zampini ierr = MatDestroy(&subdomain_adj);CHKERRQ(ierr); 4767e7931f94SStefano Zampini ierr = MatPartitioningDestroy(&partitioner);CHKERRQ(ierr); 4768e7931f94SStefano Zampini } 4769e7931f94SStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4770e7931f94SStefano Zampini 4771e7931f94SStefano Zampini /* assemble parallel IS for sends */ 4772e7931f94SStefano Zampini i = 1; 4773e7931f94SStefano Zampini if (color) i=0; 4774e7931f94SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,&ranks_send_to);CHKERRQ(ierr); 4775e7931f94SStefano Zampini /* get back IS */ 4776e7931f94SStefano Zampini *is_sends = ranks_send_to; 4777e7931f94SStefano Zampini PetscFunctionReturn(0); 4778e7931f94SStefano Zampini } 4779e7931f94SStefano Zampini 4780e7931f94SStefano Zampini typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate; 4781e7931f94SStefano Zampini 4782e7931f94SStefano Zampini #undef __FUNCT__ 4783e7931f94SStefano Zampini #define __FUNCT__ "MatISSubassemble" 478453a05cb3SStefano 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[]) 4785e7931f94SStefano Zampini { 478670cf5478SStefano Zampini Mat local_mat; 4787e7931f94SStefano Zampini IS is_sends_internal; 47889d30be91SStefano Zampini PetscInt rows,cols,new_local_rows; 478928143c3dSStefano Zampini PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals; 47909d30be91SStefano Zampini PetscBool ismatis,isdense,newisdense,destroy_mat; 4791e7931f94SStefano Zampini ISLocalToGlobalMapping l2gmap; 4792e7931f94SStefano Zampini PetscInt* l2gmap_indices; 4793e7931f94SStefano Zampini const PetscInt* is_indices; 4794e7931f94SStefano Zampini MatType new_local_type; 4795e7931f94SStefano Zampini /* buffers */ 4796e7931f94SStefano Zampini PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs; 479728143c3dSStefano Zampini PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is; 47989d30be91SStefano Zampini PetscInt *recv_buffer_idxs_local; 4799e7931f94SStefano Zampini PetscScalar *ptr_vals,*send_buffer_vals,*recv_buffer_vals; 4800e7931f94SStefano Zampini /* MPI */ 480128143c3dSStefano Zampini MPI_Comm comm,comm_n; 480228143c3dSStefano Zampini PetscSubcomm subcomm; 4803e7931f94SStefano Zampini PetscMPIInt n_sends,n_recvs,commsize; 480428143c3dSStefano Zampini PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is; 480528143c3dSStefano Zampini PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals; 480628143c3dSStefano Zampini PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,source_dest; 480728143c3dSStefano Zampini MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals; 480828143c3dSStefano Zampini MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals; 4809e7931f94SStefano Zampini PetscErrorCode ierr; 4810e7931f94SStefano Zampini 4811e7931f94SStefano Zampini PetscFunctionBegin; 481228143c3dSStefano Zampini /* TODO: add missing checks */ 481328143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,n_subdomains,3); 481428143c3dSStefano Zampini PetscValidLogicalCollectiveBool(mat,restrict_comm,4); 481528143c3dSStefano Zampini PetscValidLogicalCollectiveEnum(mat,reuse,5); 481628143c3dSStefano Zampini PetscValidLogicalCollectiveInt(mat,nis,7); 4817e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);CHKERRQ(ierr); 481828143c3dSStefano Zampini if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",__FUNCT__); 4819e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 4820e7931f94SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);CHKERRQ(ierr); 4821e7931f94SStefano Zampini if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE"); 4822e7931f94SStefano Zampini ierr = MatGetSize(local_mat,&rows,&cols);CHKERRQ(ierr); 4823e7931f94SStefano Zampini if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square"); 482428143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX && *mat_n) { 482570cf5478SStefano Zampini PetscInt mrows,mcols,mnrows,mncols; 482670cf5478SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);CHKERRQ(ierr); 482728143c3dSStefano Zampini if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS"); 482870cf5478SStefano Zampini ierr = MatGetSize(mat,&mrows,&mcols);CHKERRQ(ierr); 482970cf5478SStefano Zampini ierr = MatGetSize(*mat_n,&mnrows,&mncols);CHKERRQ(ierr); 483070cf5478SStefano Zampini if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows); 483170cf5478SStefano Zampini if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols); 483270cf5478SStefano Zampini } 4833e7931f94SStefano Zampini ierr = MatGetBlockSize(local_mat,&bs);CHKERRQ(ierr); 4834e7931f94SStefano Zampini PetscValidLogicalCollectiveInt(mat,bs,0); 4835e7931f94SStefano Zampini /* prepare IS for sending if not provided */ 4836e7931f94SStefano Zampini if (!is_sends) { 483728143c3dSStefano Zampini if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains"); 4838b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(mat,n_subdomains,0,&is_sends_internal);CHKERRQ(ierr); 4839c8587f34SStefano Zampini } else { 4840e7931f94SStefano Zampini ierr = PetscObjectReference((PetscObject)is_sends);CHKERRQ(ierr); 4841e7931f94SStefano Zampini is_sends_internal = is_sends; 4842c8587f34SStefano Zampini } 4843e7931f94SStefano Zampini 4844e7931f94SStefano Zampini /* get comm */ 4845a316fed8SStefano Zampini ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 4846e7931f94SStefano Zampini 4847e7931f94SStefano Zampini /* compute number of sends */ 4848e7931f94SStefano Zampini ierr = ISGetLocalSize(is_sends_internal,&i);CHKERRQ(ierr); 4849e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&n_sends);CHKERRQ(ierr); 4850e7931f94SStefano Zampini 4851e7931f94SStefano Zampini /* compute number of receives */ 4852e7931f94SStefano Zampini ierr = MPI_Comm_size(comm,&commsize);CHKERRQ(ierr); 4853785e854fSJed Brown ierr = PetscMalloc1(commsize,&iflags);CHKERRQ(ierr); 4854e7931f94SStefano Zampini ierr = PetscMemzero(iflags,commsize*sizeof(*iflags));CHKERRQ(ierr); 4855e7931f94SStefano Zampini ierr = ISGetIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 4856e7931f94SStefano Zampini for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1; 4857e7931f94SStefano Zampini ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);CHKERRQ(ierr); 4858e7931f94SStefano Zampini ierr = PetscFree(iflags);CHKERRQ(ierr); 4859e7931f94SStefano Zampini 486028143c3dSStefano Zampini /* restrict comm if requested */ 486128143c3dSStefano Zampini subcomm = 0; 486228143c3dSStefano Zampini destroy_mat = PETSC_FALSE; 486328143c3dSStefano Zampini if (restrict_comm) { 4864779c1cceSStefano Zampini PetscMPIInt color,subcommsize; 4865779c1cceSStefano Zampini 486628143c3dSStefano Zampini color = 0; 486753a05cb3SStefano Zampini if (restrict_full) { 486853a05cb3SStefano Zampini if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */ 486953a05cb3SStefano Zampini } else { 487053a05cb3SStefano Zampini if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */ 487153a05cb3SStefano Zampini } 487228143c3dSStefano Zampini ierr = MPI_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 487328143c3dSStefano Zampini subcommsize = commsize - subcommsize; 487428143c3dSStefano Zampini /* check if reuse has been requested */ 487528143c3dSStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 487628143c3dSStefano Zampini if (*mat_n) { 487728143c3dSStefano Zampini PetscMPIInt subcommsize2; 487828143c3dSStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);CHKERRQ(ierr); 487928143c3dSStefano Zampini if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2); 488028143c3dSStefano Zampini comm_n = PetscObjectComm((PetscObject)*mat_n); 488128143c3dSStefano Zampini } else { 488228143c3dSStefano Zampini comm_n = PETSC_COMM_SELF; 488328143c3dSStefano Zampini } 488428143c3dSStefano Zampini } else { /* MAT_INITIAL_MATRIX */ 4885779c1cceSStefano Zampini PetscMPIInt rank; 4886779c1cceSStefano Zampini 4887779c1cceSStefano Zampini ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 488828143c3dSStefano Zampini ierr = PetscSubcommCreate(comm,&subcomm);CHKERRQ(ierr); 488928143c3dSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 489028143c3dSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4891306c2d5bSBarry Smith comm_n = PetscSubcommChild(subcomm); 489228143c3dSStefano Zampini } 489328143c3dSStefano Zampini /* flag to destroy *mat_n if not significative */ 489428143c3dSStefano Zampini if (color) destroy_mat = PETSC_TRUE; 489528143c3dSStefano Zampini } else { 489628143c3dSStefano Zampini comm_n = comm; 489728143c3dSStefano Zampini } 489828143c3dSStefano Zampini 4899e7931f94SStefano Zampini /* prepare send/receive buffers */ 4900785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_idxs);CHKERRQ(ierr); 4901e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_idxs,commsize*sizeof(*ilengths_idxs));CHKERRQ(ierr); 4902785e854fSJed Brown ierr = PetscMalloc1(commsize,&ilengths_vals);CHKERRQ(ierr); 4903e7931f94SStefano Zampini ierr = PetscMemzero(ilengths_vals,commsize*sizeof(*ilengths_vals));CHKERRQ(ierr); 490428143c3dSStefano Zampini if (nis) { 4905854ce69bSBarry Smith ierr = PetscCalloc1(commsize,&ilengths_idxs_is);CHKERRQ(ierr); 490628143c3dSStefano Zampini } 4907e7931f94SStefano Zampini 490828143c3dSStefano Zampini /* Get data from local matrices */ 4909e7931f94SStefano Zampini if (!isdense) { 4910a26c9d0eSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented"); 4911e7931f94SStefano Zampini /* TODO: See below some guidelines on how to prepare the local buffers */ 4912e7931f94SStefano Zampini /* 4913e7931f94SStefano Zampini send_buffer_vals should contain the raw values of the local matrix 4914e7931f94SStefano Zampini send_buffer_idxs should contain: 4915e7931f94SStefano Zampini - MatType_PRIVATE type 4916e7931f94SStefano Zampini - PetscInt size_of_l2gmap 4917e7931f94SStefano Zampini - PetscInt global_row_indices[size_of_l2gmap] 4918e7931f94SStefano Zampini - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values 4919e7931f94SStefano Zampini */ 4920e7931f94SStefano Zampini } else { 4921e7931f94SStefano Zampini ierr = MatDenseGetArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 49223bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);CHKERRQ(ierr); 4923854ce69bSBarry Smith ierr = PetscMalloc1(i+2,&send_buffer_idxs);CHKERRQ(ierr); 4924e7931f94SStefano Zampini send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE; 4925e7931f94SStefano Zampini send_buffer_idxs[1] = i; 49263bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4927e7931f94SStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs[2],ptr_idxs,i*sizeof(PetscInt));CHKERRQ(ierr); 49283bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);CHKERRQ(ierr); 4929e7931f94SStefano Zampini ierr = PetscMPIIntCast(i,&len);CHKERRQ(ierr); 4930e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 4931e7931f94SStefano Zampini ilengths_vals[is_indices[i]] = len*len; 4932e7931f94SStefano Zampini ilengths_idxs[is_indices[i]] = len+2; 4933c8587f34SStefano Zampini } 4934c8587f34SStefano Zampini } 4935e7931f94SStefano Zampini ierr = PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);CHKERRQ(ierr); 493628143c3dSStefano Zampini /* additional is (if any) */ 493728143c3dSStefano Zampini if (nis) { 493828143c3dSStefano Zampini PetscMPIInt psum; 493928143c3dSStefano Zampini PetscInt j; 494028143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 494128143c3dSStefano Zampini PetscInt plen; 494228143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 494328143c3dSStefano Zampini ierr = PetscMPIIntCast(plen,&len);CHKERRQ(ierr); 494428143c3dSStefano Zampini psum += len+1; /* indices + lenght */ 494528143c3dSStefano Zampini } 4946854ce69bSBarry Smith ierr = PetscMalloc1(psum,&send_buffer_idxs_is);CHKERRQ(ierr); 494728143c3dSStefano Zampini for (j=0,psum=0;j<nis;j++) { 494828143c3dSStefano Zampini PetscInt plen; 494928143c3dSStefano Zampini const PetscInt *is_array_idxs; 495028143c3dSStefano Zampini ierr = ISGetLocalSize(isarray[j],&plen);CHKERRQ(ierr); 495128143c3dSStefano Zampini send_buffer_idxs_is[psum] = plen; 495228143c3dSStefano Zampini ierr = ISGetIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 495328143c3dSStefano Zampini ierr = PetscMemcpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen*sizeof(PetscInt));CHKERRQ(ierr); 495428143c3dSStefano Zampini ierr = ISRestoreIndices(isarray[j],&is_array_idxs);CHKERRQ(ierr); 495528143c3dSStefano Zampini psum += plen+1; /* indices + lenght */ 495628143c3dSStefano Zampini } 495728143c3dSStefano Zampini for (i=0;i<n_sends;i++) { 495828143c3dSStefano Zampini ilengths_idxs_is[is_indices[i]] = psum; 495928143c3dSStefano Zampini } 496028143c3dSStefano Zampini ierr = PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);CHKERRQ(ierr); 496128143c3dSStefano Zampini } 496228143c3dSStefano Zampini 4963e7931f94SStefano Zampini buf_size_idxs = 0; 4964e7931f94SStefano Zampini buf_size_vals = 0; 496528143c3dSStefano Zampini buf_size_idxs_is = 0; 4966e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4967e7931f94SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 4968e7931f94SStefano Zampini buf_size_vals += (PetscInt)olengths_vals[i]; 496928143c3dSStefano Zampini if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i]; 4970e7931f94SStefano Zampini } 4971785e854fSJed Brown ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);CHKERRQ(ierr); 4972785e854fSJed Brown ierr = PetscMalloc1(buf_size_vals,&recv_buffer_vals);CHKERRQ(ierr); 497395ecbf38SStefano Zampini ierr = PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);CHKERRQ(ierr); 4974e7931f94SStefano Zampini 4975e7931f94SStefano Zampini /* get new tags for clean communications */ 4976e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);CHKERRQ(ierr); 4977e7931f94SStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_vals);CHKERRQ(ierr); 497828143c3dSStefano Zampini ierr = PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);CHKERRQ(ierr); 4979e7931f94SStefano Zampini 4980e7931f94SStefano Zampini /* allocate for requests */ 4981785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_idxs);CHKERRQ(ierr); 4982785e854fSJed Brown ierr = PetscMalloc1(n_sends,&send_req_vals);CHKERRQ(ierr); 498395ecbf38SStefano Zampini ierr = PetscMalloc1(n_sends,&send_req_idxs_is);CHKERRQ(ierr); 4984785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_idxs);CHKERRQ(ierr); 4985785e854fSJed Brown ierr = PetscMalloc1(n_recvs,&recv_req_vals);CHKERRQ(ierr); 498695ecbf38SStefano Zampini ierr = PetscMalloc1(n_recvs,&recv_req_idxs_is);CHKERRQ(ierr); 4987e7931f94SStefano Zampini 4988e7931f94SStefano Zampini /* communications */ 4989e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 4990e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 499128143c3dSStefano Zampini ptr_idxs_is = recv_buffer_idxs_is; 4992e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 4993e7931f94SStefano Zampini source_dest = onodes[i]; 4994e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);CHKERRQ(ierr); 4995e7931f94SStefano Zampini ierr = MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);CHKERRQ(ierr); 4996e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 4997e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 499828143c3dSStefano Zampini if (nis) { 499928143c3dSStefano 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); 500028143c3dSStefano Zampini ptr_idxs_is += olengths_idxs_is[i]; 500128143c3dSStefano Zampini } 5002e7931f94SStefano Zampini } 5003e7931f94SStefano Zampini for (i=0;i<n_sends;i++) { 5004e7931f94SStefano Zampini ierr = PetscMPIIntCast(is_indices[i],&source_dest);CHKERRQ(ierr); 5005e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);CHKERRQ(ierr); 5006e7931f94SStefano Zampini ierr = MPI_Isend(send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);CHKERRQ(ierr); 500728143c3dSStefano Zampini if (nis) { 500828143c3dSStefano 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); 500928143c3dSStefano Zampini } 5010e7931f94SStefano Zampini } 5011e7931f94SStefano Zampini ierr = ISRestoreIndices(is_sends_internal,&is_indices);CHKERRQ(ierr); 5012e7931f94SStefano Zampini ierr = ISDestroy(&is_sends_internal);CHKERRQ(ierr); 5013e7931f94SStefano Zampini 5014e7931f94SStefano Zampini /* assemble new l2g map */ 5015e7931f94SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5016e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50179d30be91SStefano Zampini new_local_rows = 0; 5018e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50199d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5020e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5021e7931f94SStefano Zampini } 50229d30be91SStefano Zampini ierr = PetscMalloc1(new_local_rows,&l2gmap_indices);CHKERRQ(ierr); 5023e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 50249d30be91SStefano Zampini new_local_rows = 0; 5025e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 50269d30be91SStefano Zampini ierr = PetscMemcpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,(*(ptr_idxs+1))*sizeof(PetscInt));CHKERRQ(ierr); 50279d30be91SStefano Zampini new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */ 5028e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5029e7931f94SStefano Zampini } 50309d30be91SStefano Zampini ierr = PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);CHKERRQ(ierr); 50319d30be91SStefano Zampini ierr = ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);CHKERRQ(ierr); 5032e7931f94SStefano Zampini ierr = PetscFree(l2gmap_indices);CHKERRQ(ierr); 5033e7931f94SStefano Zampini 5034e7931f94SStefano Zampini /* infer new local matrix type from received local matrices type */ 5035e7931f94SStefano Zampini /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */ 5036e7931f94SStefano 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) */ 5037e7931f94SStefano Zampini if (n_recvs) { 503828143c3dSStefano Zampini MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0]; 5039e7931f94SStefano Zampini ptr_idxs = recv_buffer_idxs; 5040e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5041e7931f94SStefano Zampini if ((PetscInt)new_local_type_private != *ptr_idxs) { 5042e7931f94SStefano Zampini new_local_type_private = MATAIJ_PRIVATE; 5043e7931f94SStefano Zampini break; 5044e7931f94SStefano Zampini } 5045e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5046e7931f94SStefano Zampini } 5047e7931f94SStefano Zampini switch (new_local_type_private) { 504828143c3dSStefano Zampini case MATDENSE_PRIVATE: 504928143c3dSStefano Zampini if (n_recvs>1) { /* subassembling of dense matrices does not give a dense matrix! */ 5050e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5051e7931f94SStefano Zampini bs = 1; 505228143c3dSStefano Zampini } else { /* if I receive only 1 dense matrix */ 505328143c3dSStefano Zampini new_local_type = MATSEQDENSE; 505428143c3dSStefano Zampini bs = 1; 505528143c3dSStefano Zampini } 5056e7931f94SStefano Zampini break; 5057e7931f94SStefano Zampini case MATAIJ_PRIVATE: 5058e7931f94SStefano Zampini new_local_type = MATSEQAIJ; 5059e7931f94SStefano Zampini bs = 1; 5060e7931f94SStefano Zampini break; 5061e7931f94SStefano Zampini case MATBAIJ_PRIVATE: 5062e7931f94SStefano Zampini new_local_type = MATSEQBAIJ; 5063e7931f94SStefano Zampini break; 5064e7931f94SStefano Zampini case MATSBAIJ_PRIVATE: 5065e7931f94SStefano Zampini new_local_type = MATSEQSBAIJ; 5066e7931f94SStefano Zampini break; 5067e7931f94SStefano Zampini default: 50689d30be91SStefano Zampini SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,__FUNCT__); 5069e7931f94SStefano Zampini break; 5070e7931f94SStefano Zampini } 507128143c3dSStefano Zampini } else { /* by default, new_local_type is seqdense */ 507228143c3dSStefano Zampini new_local_type = MATSEQDENSE; 507328143c3dSStefano Zampini bs = 1; 5074e7931f94SStefano Zampini } 5075e7931f94SStefano Zampini 507670cf5478SStefano Zampini /* create MATIS object if needed */ 507770cf5478SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 5078e7931f94SStefano Zampini ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 5079e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);CHKERRQ(ierr); 508070cf5478SStefano Zampini } else { 508170cf5478SStefano Zampini /* it also destroys the local matrices */ 508270cf5478SStefano Zampini ierr = MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);CHKERRQ(ierr); 508370cf5478SStefano Zampini } 508470cf5478SStefano Zampini ierr = MatISGetLocalMat(*mat_n,&local_mat);CHKERRQ(ierr); 5085e7931f94SStefano Zampini ierr = MatSetType(local_mat,new_local_type);CHKERRQ(ierr); 50869d30be91SStefano Zampini 50879d30be91SStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 50889d30be91SStefano Zampini 50899d30be91SStefano Zampini /* Global to local map of received indices */ 50909d30be91SStefano Zampini ierr = PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local);CHKERRQ(ierr); /* needed for values insertion */ 50919d30be91SStefano Zampini ierr = ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);CHKERRQ(ierr); 50929d30be91SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap);CHKERRQ(ierr); 50939d30be91SStefano Zampini 50949d30be91SStefano Zampini /* restore attributes -> type of incoming data and its size */ 50959d30be91SStefano Zampini buf_size_idxs = 0; 50969d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 50979d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs]; 50989d30be91SStefano Zampini recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1]; 50999d30be91SStefano Zampini buf_size_idxs += (PetscInt)olengths_idxs[i]; 51009d30be91SStefano Zampini } 51019d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs);CHKERRQ(ierr); 51029d30be91SStefano Zampini 51039d30be91SStefano Zampini /* set preallocation */ 51049d30be91SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);CHKERRQ(ierr); 51059d30be91SStefano Zampini if (!newisdense) { 51069d30be91SStefano Zampini PetscInt *new_local_nnz=0; 51079d30be91SStefano Zampini 51089d30be91SStefano Zampini ptr_vals = recv_buffer_vals; 51099d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 51109d30be91SStefano Zampini if (n_recvs) { 51119d30be91SStefano Zampini ierr = PetscCalloc1(new_local_rows,&new_local_nnz);CHKERRQ(ierr); 51129d30be91SStefano Zampini } 51139d30be91SStefano Zampini for (i=0;i<n_recvs;i++) { 51149d30be91SStefano Zampini PetscInt j; 51159d30be91SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */ 51169d30be91SStefano Zampini for (j=0;j<*(ptr_idxs+1);j++) { 51179d30be91SStefano Zampini new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1); 51189d30be91SStefano Zampini } 51199d30be91SStefano Zampini } else { 51209d30be91SStefano Zampini /* TODO */ 51219d30be91SStefano Zampini } 51229d30be91SStefano Zampini ptr_idxs += olengths_idxs[i]; 51239d30be91SStefano Zampini } 51249d30be91SStefano Zampini if (new_local_nnz) { 51259d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows); 51269d30be91SStefano Zampini ierr = MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);CHKERRQ(ierr); 51279d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs; 51289d30be91SStefano Zampini ierr = MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51299d30be91SStefano Zampini for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0); 51309d30be91SStefano Zampini ierr = MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);CHKERRQ(ierr); 51319d30be91SStefano Zampini } else { 51329d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51339d30be91SStefano Zampini } 51349d30be91SStefano Zampini ierr = PetscFree(new_local_nnz);CHKERRQ(ierr); 51359d30be91SStefano Zampini } else { 51369d30be91SStefano Zampini ierr = MatSetUp(local_mat);CHKERRQ(ierr); 51379d30be91SStefano Zampini } 5138e7931f94SStefano Zampini 5139e7931f94SStefano Zampini /* set values */ 5140e7931f94SStefano Zampini ptr_vals = recv_buffer_vals; 51419d30be91SStefano Zampini ptr_idxs = recv_buffer_idxs_local; 5142e7931f94SStefano Zampini for (i=0;i<n_recvs;i++) { 5143e7931f94SStefano Zampini if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */ 5144e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr); 51459d30be91SStefano Zampini ierr = MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);CHKERRQ(ierr); 5146e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5147e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);CHKERRQ(ierr); 5148e7931f94SStefano Zampini ierr = MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr); 514928143c3dSStefano Zampini } else { 515028143c3dSStefano Zampini /* TODO */ 5151e7931f94SStefano Zampini } 5152e7931f94SStefano Zampini ptr_idxs += olengths_idxs[i]; 5153e7931f94SStefano Zampini ptr_vals += olengths_vals[i]; 5154e7931f94SStefano Zampini } 5155e7931f94SStefano Zampini ierr = MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5156e7931f94SStefano Zampini ierr = MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 515770cf5478SStefano Zampini ierr = MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 515870cf5478SStefano Zampini ierr = MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 51599d30be91SStefano Zampini ierr = PetscFree(recv_buffer_vals);CHKERRQ(ierr); 51609d30be91SStefano Zampini ierr = PetscFree(recv_buffer_idxs_local);CHKERRQ(ierr); 5161e7931f94SStefano Zampini 5162dfd14d43SStefano Zampini #if 0 516328143c3dSStefano Zampini if (!restrict_comm) { /* check */ 5164e7931f94SStefano Zampini Vec lvec,rvec; 5165e7931f94SStefano Zampini PetscReal infty_error; 5166e7931f94SStefano Zampini 51672a7a6963SBarry Smith ierr = MatCreateVecs(mat,&rvec,&lvec);CHKERRQ(ierr); 5168e7931f94SStefano Zampini ierr = VecSetRandom(rvec,NULL);CHKERRQ(ierr); 5169e7931f94SStefano Zampini ierr = MatMult(mat,rvec,lvec);CHKERRQ(ierr); 5170e7931f94SStefano Zampini ierr = VecScale(lvec,-1.0);CHKERRQ(ierr); 517170cf5478SStefano Zampini ierr = MatMultAdd(*mat_n,rvec,lvec,lvec);CHKERRQ(ierr); 5172e7931f94SStefano Zampini ierr = VecNorm(lvec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 5173e7931f94SStefano Zampini ierr = PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error); 5174e7931f94SStefano Zampini ierr = VecDestroy(&rvec);CHKERRQ(ierr); 5175e7931f94SStefano Zampini ierr = VecDestroy(&lvec);CHKERRQ(ierr); 5176e7931f94SStefano Zampini } 517728143c3dSStefano Zampini #endif 5178e7931f94SStefano Zampini 517928143c3dSStefano Zampini /* assemble new additional is (if any) */ 518028143c3dSStefano Zampini if (nis) { 518128143c3dSStefano Zampini PetscInt **temp_idxs,*count_is,j,psum; 518228143c3dSStefano Zampini 518328143c3dSStefano Zampini ierr = MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5184854ce69bSBarry Smith ierr = PetscCalloc1(nis,&count_is);CHKERRQ(ierr); 518528143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 518628143c3dSStefano Zampini psum = 0; 518728143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 518828143c3dSStefano Zampini for (j=0;j<nis;j++) { 518928143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 519028143c3dSStefano Zampini count_is[j] += plen; /* increment counting of buffer for j-th IS */ 519128143c3dSStefano Zampini psum += plen; 519228143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 519328143c3dSStefano Zampini } 519428143c3dSStefano Zampini } 5195854ce69bSBarry Smith ierr = PetscMalloc1(nis,&temp_idxs);CHKERRQ(ierr); 5196854ce69bSBarry Smith ierr = PetscMalloc1(psum,&temp_idxs[0]);CHKERRQ(ierr); 519728143c3dSStefano Zampini for (i=1;i<nis;i++) { 519828143c3dSStefano Zampini temp_idxs[i] = temp_idxs[i-1]+count_is[i-1]; 519928143c3dSStefano Zampini } 520028143c3dSStefano Zampini ierr = PetscMemzero(count_is,nis*sizeof(PetscInt));CHKERRQ(ierr); 520128143c3dSStefano Zampini ptr_idxs = recv_buffer_idxs_is; 520228143c3dSStefano Zampini for (i=0;i<n_recvs;i++) { 520328143c3dSStefano Zampini for (j=0;j<nis;j++) { 520428143c3dSStefano Zampini PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */ 520528143c3dSStefano Zampini ierr = PetscMemcpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen*sizeof(PetscInt));CHKERRQ(ierr); 520628143c3dSStefano Zampini count_is[j] += plen; /* increment starting point of buffer for j-th IS */ 520728143c3dSStefano Zampini ptr_idxs += plen+1; /* shift pointer to received data */ 520828143c3dSStefano Zampini } 520928143c3dSStefano Zampini } 521028143c3dSStefano Zampini for (i=0;i<nis;i++) { 521128143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 521228143c3dSStefano Zampini ierr = PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);CHKERRQ(ierr);CHKERRQ(ierr); 521328143c3dSStefano Zampini ierr = ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 521428143c3dSStefano Zampini } 521528143c3dSStefano Zampini ierr = PetscFree(count_is);CHKERRQ(ierr); 521628143c3dSStefano Zampini ierr = PetscFree(temp_idxs[0]);CHKERRQ(ierr); 521728143c3dSStefano Zampini ierr = PetscFree(temp_idxs);CHKERRQ(ierr); 521828143c3dSStefano Zampini } 5219e7931f94SStefano Zampini /* free workspace */ 522028143c3dSStefano Zampini ierr = PetscFree(recv_buffer_idxs_is);CHKERRQ(ierr); 5221e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5222e7931f94SStefano Zampini ierr = PetscFree(send_buffer_idxs);CHKERRQ(ierr); 5223e7931f94SStefano Zampini ierr = MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 5224e7931f94SStefano Zampini if (isdense) { 5225e7931f94SStefano Zampini ierr = MatISGetLocalMat(mat,&local_mat);CHKERRQ(ierr); 5226e7931f94SStefano Zampini ierr = MatDenseRestoreArray(local_mat,&send_buffer_vals);CHKERRQ(ierr); 5227e7931f94SStefano Zampini } else { 5228e7931f94SStefano Zampini /* ierr = PetscFree(send_buffer_vals);CHKERRQ(ierr); */ 5229e7931f94SStefano Zampini } 523028143c3dSStefano Zampini if (nis) { 523128143c3dSStefano Zampini ierr = MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);CHKERRQ(ierr); 523228143c3dSStefano Zampini ierr = PetscFree(send_buffer_idxs_is);CHKERRQ(ierr); 523328143c3dSStefano Zampini } 5234e7931f94SStefano Zampini ierr = PetscFree(recv_req_idxs);CHKERRQ(ierr); 5235e7931f94SStefano Zampini ierr = PetscFree(recv_req_vals);CHKERRQ(ierr); 523628143c3dSStefano Zampini ierr = PetscFree(recv_req_idxs_is);CHKERRQ(ierr); 5237e7931f94SStefano Zampini ierr = PetscFree(send_req_idxs);CHKERRQ(ierr); 5238e7931f94SStefano Zampini ierr = PetscFree(send_req_vals);CHKERRQ(ierr); 523928143c3dSStefano Zampini ierr = PetscFree(send_req_idxs_is);CHKERRQ(ierr); 5240e7931f94SStefano Zampini ierr = PetscFree(ilengths_vals);CHKERRQ(ierr); 5241e7931f94SStefano Zampini ierr = PetscFree(ilengths_idxs);CHKERRQ(ierr); 5242e7931f94SStefano Zampini ierr = PetscFree(olengths_vals);CHKERRQ(ierr); 5243e7931f94SStefano Zampini ierr = PetscFree(olengths_idxs);CHKERRQ(ierr); 5244e7931f94SStefano Zampini ierr = PetscFree(onodes);CHKERRQ(ierr); 524528143c3dSStefano Zampini if (nis) { 524628143c3dSStefano Zampini ierr = PetscFree(ilengths_idxs_is);CHKERRQ(ierr); 524728143c3dSStefano Zampini ierr = PetscFree(olengths_idxs_is);CHKERRQ(ierr); 524828143c3dSStefano Zampini ierr = PetscFree(onodes_is);CHKERRQ(ierr); 524928143c3dSStefano Zampini } 525028143c3dSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 525128143c3dSStefano Zampini if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */ 525228143c3dSStefano Zampini ierr = MatDestroy(mat_n);CHKERRQ(ierr); 525328143c3dSStefano Zampini for (i=0;i<nis;i++) { 525428143c3dSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 525528143c3dSStefano Zampini } 525653a05cb3SStefano Zampini *mat_n = NULL; 525728143c3dSStefano Zampini } 5258e7931f94SStefano Zampini PetscFunctionReturn(0); 5259e7931f94SStefano Zampini } 5260a57a6d2fSStefano Zampini 526112edc857SStefano Zampini /* temporary hack into ksp private data structure */ 5262af0996ceSBarry Smith #include <petsc/private/kspimpl.h> 526312edc857SStefano Zampini 5264c8587f34SStefano Zampini #undef __FUNCT__ 5265c8587f34SStefano Zampini #define __FUNCT__ "PCBDDCSetUpCoarseSolver" 5266c8587f34SStefano Zampini PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals) 5267c8587f34SStefano Zampini { 5268c8587f34SStefano Zampini PC_BDDC *pcbddc = (PC_BDDC*)pc->data; 5269c8587f34SStefano Zampini PC_IS *pcis = (PC_IS*)pc->data; 527020a2ab83SStefano Zampini Mat coarse_mat,coarse_mat_is,coarse_submat_dense; 52719881197aSStefano Zampini MatNullSpace CoarseNullSpace=NULL; 527220a2ab83SStefano Zampini ISLocalToGlobalMapping coarse_islg; 52736e683305SStefano Zampini IS coarse_is,*isarray; 52746e683305SStefano Zampini PetscInt i,im_active=-1,active_procs=-1; 527530368db7SStefano Zampini PetscInt nis,nisdofs,nisneu,nisvert; 5276f9eb5b7dSStefano Zampini PC pc_temp; 5277c8587f34SStefano Zampini PCType coarse_pc_type; 5278c8587f34SStefano Zampini KSPType coarse_ksp_type; 5279f9eb5b7dSStefano Zampini PetscBool multilevel_requested,multilevel_allowed; 52804f3a063dSStefano Zampini PetscBool isredundant,isbddc,isnn,coarse_reuse; 52816e683305SStefano Zampini Mat t_coarse_mat_is; 52826e683305SStefano Zampini PetscInt void_procs,ncoarse_ml,ncoarse_ds,ncoarse; 52836e683305SStefano Zampini PetscMPIInt all_procs; 528474e2c79eSStefano Zampini PetscBool csin_ml,csin_ds,csin,csin_type_simple,redist; 528568457ee5SStefano Zampini PetscBool compute_vecs = PETSC_FALSE; 528622bc73bbSStefano Zampini PetscScalar *array; 52879881197aSStefano Zampini PetscErrorCode ierr; 5288fdc09c96SStefano Zampini 5289c8587f34SStefano Zampini PetscFunctionBegin; 5290c8587f34SStefano Zampini /* Assign global numbering to coarse dofs */ 529168457ee5SStefano 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 */ 5292fa7f1dd8SStefano Zampini PetscInt ocoarse_size; 52935a75c04eSSatish Balay compute_vecs = PETSC_TRUE; 5294fa7f1dd8SStefano Zampini ocoarse_size = pcbddc->coarse_size; 5295f4ddd8eeSStefano Zampini ierr = PetscFree(pcbddc->global_primal_indices);CHKERRQ(ierr); 5296f4ddd8eeSStefano Zampini ierr = PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);CHKERRQ(ierr); 5297f4ddd8eeSStefano Zampini /* see if we can avoid some work */ 5298fa7f1dd8SStefano Zampini if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */ 529951bea450SStefano Zampini /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */ 530051bea450SStefano Zampini if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) { 5301dc4bcba2SStefano Zampini PC pc; 5302dc4bcba2SStefano Zampini PetscBool isbddc; 5303dc4bcba2SStefano Zampini 5304dc4bcba2SStefano Zampini /* temporary workaround since PCBDDC does not have a reset method so far */ 5305dc4bcba2SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc);CHKERRQ(ierr); 5306dc4bcba2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc,PCBDDC,&isbddc);CHKERRQ(ierr); 5307dc4bcba2SStefano Zampini if (isbddc) { 530863c961adSStefano Zampini ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr); 530963c961adSStefano Zampini } else { 5310727cdba6SStefano Zampini ierr = KSPReset(pcbddc->coarse_ksp);CHKERRQ(ierr); 531163c961adSStefano Zampini } 5312fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5313fa7f1dd8SStefano Zampini } else { /* we can safely reuse already computed coarse matrix */ 5314fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5315f4ddd8eeSStefano Zampini } 5316fa7f1dd8SStefano Zampini } else { /* there's no coarse ksp, so we need to create the coarse matrix too */ 5317fa7f1dd8SStefano Zampini coarse_reuse = PETSC_FALSE; 5318f4ddd8eeSStefano Zampini } 531970cf5478SStefano Zampini /* reset any subassembling information */ 532070cf5478SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling);CHKERRQ(ierr); 53216e683305SStefano Zampini ierr = ISDestroy(&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 53226e683305SStefano Zampini } else { /* primal space is unchanged, so we can reuse coarse matrix */ 5323fa7f1dd8SStefano Zampini coarse_reuse = PETSC_TRUE; 5324f4ddd8eeSStefano Zampini } 5325c8587f34SStefano Zampini 53266e683305SStefano Zampini /* count "active" (i.e. with positive local size) and "void" processes */ 53272b510759SStefano Zampini im_active = !!(pcis->n); 53282b510759SStefano Zampini ierr = MPI_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 53296e683305SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)pc),&all_procs);CHKERRQ(ierr); 53306e683305SStefano Zampini void_procs = all_procs-active_procs; 53316e683305SStefano Zampini csin_type_simple = PETSC_TRUE; 533274e2c79eSStefano Zampini redist = PETSC_FALSE; 533322bc73bbSStefano Zampini if (pcbddc->current_level && void_procs) { 53346e683305SStefano Zampini csin_ml = PETSC_TRUE; 53356e683305SStefano Zampini ncoarse_ml = void_procs; 5336779c1cceSStefano Zampini /* it has no sense to redistribute on a set of processors larger than the number of active processes */ 5337779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < active_procs) { 53386e683305SStefano Zampini csin_ds = PETSC_TRUE; 533918a45a71SStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 534018a45a71SStefano Zampini redist = PETSC_TRUE; 534118a45a71SStefano Zampini } else { 53426e683305SStefano Zampini csin_ds = PETSC_TRUE; 5343779c1cceSStefano Zampini ncoarse_ds = active_procs; 5344779c1cceSStefano Zampini redist = PETSC_TRUE; 534518a45a71SStefano Zampini } 53466e683305SStefano Zampini } else { 53476e683305SStefano Zampini csin_ml = PETSC_FALSE; 53486e683305SStefano Zampini ncoarse_ml = all_procs; 53496e683305SStefano Zampini if (void_procs) { 53506e683305SStefano Zampini csin_ds = PETSC_TRUE; 53516e683305SStefano Zampini ncoarse_ds = void_procs; 53526e683305SStefano Zampini csin_type_simple = PETSC_FALSE; 53536e683305SStefano Zampini } else { 5354779c1cceSStefano Zampini if (pcbddc->redistribute_coarse > 0 && pcbddc->redistribute_coarse < all_procs) { 535574e2c79eSStefano Zampini csin_ds = PETSC_TRUE; 535674e2c79eSStefano Zampini ncoarse_ds = pcbddc->redistribute_coarse; 535774e2c79eSStefano Zampini redist = PETSC_TRUE; 535874e2c79eSStefano Zampini } else { 53596e683305SStefano Zampini csin_ds = PETSC_FALSE; 53606e683305SStefano Zampini ncoarse_ds = all_procs; 53616e683305SStefano Zampini } 53626e683305SStefano Zampini } 536374e2c79eSStefano Zampini } 53646e683305SStefano Zampini 53656e683305SStefano Zampini /* 53666e683305SStefano Zampini test if we can go multilevel: three conditions must be satisfied: 53676e683305SStefano Zampini - we have not exceeded the number of levels requested 53686e683305SStefano Zampini - we can actually subassemble the active processes 53696e683305SStefano Zampini - we can find a suitable number of MPI processes where we can place the subassembled problem 53706e683305SStefano Zampini */ 53716e683305SStefano Zampini multilevel_allowed = PETSC_FALSE; 53726e683305SStefano Zampini multilevel_requested = PETSC_FALSE; 53736e683305SStefano Zampini if (pcbddc->current_level < pcbddc->max_levels) { 53746e683305SStefano Zampini multilevel_requested = PETSC_TRUE; 53756e683305SStefano Zampini if (active_procs/pcbddc->coarsening_ratio < 2 || ncoarse_ml/pcbddc->coarsening_ratio < 2) { 5376f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_FALSE; 53772b510759SStefano Zampini } else { 5378f9eb5b7dSStefano Zampini multilevel_allowed = PETSC_TRUE; 5379c8587f34SStefano Zampini } 5380c8587f34SStefano Zampini } 53816e683305SStefano Zampini /* determine number of process partecipating to coarse solver */ 53826e683305SStefano Zampini if (multilevel_allowed) { 53836e683305SStefano Zampini ncoarse = ncoarse_ml; 53846e683305SStefano Zampini csin = csin_ml; 538558da7f69SStefano Zampini redist = PETSC_FALSE; 53866e683305SStefano Zampini } else { 53876e683305SStefano Zampini ncoarse = ncoarse_ds; 53886e683305SStefano Zampini csin = csin_ds; 53896e683305SStefano Zampini } 5390e7931f94SStefano Zampini 5391abbbba34SStefano Zampini /* creates temporary l2gmap and IS for coarse indexes */ 5392abbbba34SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);CHKERRQ(ierr); 5393abbbba34SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);CHKERRQ(ierr); 5394abbbba34SStefano Zampini 5395abbbba34SStefano Zampini /* creates temporary MATIS object for coarse matrix */ 539622bc73bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,NULL,&coarse_submat_dense);CHKERRQ(ierr); 539722bc73bbSStefano Zampini ierr = MatDenseGetArray(coarse_submat_dense,&array);CHKERRQ(ierr); 539822bc73bbSStefano Zampini ierr = PetscMemcpy(array,coarse_submat_vals,sizeof(*coarse_submat_vals)*pcbddc->local_primal_size*pcbddc->local_primal_size);CHKERRQ(ierr); 539922bc73bbSStefano Zampini ierr = MatDenseRestoreArray(coarse_submat_dense,&array);CHKERRQ(ierr); 5400e176bc59SStefano 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); 54016e683305SStefano Zampini ierr = MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);CHKERRQ(ierr); 54026e683305SStefano Zampini ierr = MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 54036e683305SStefano Zampini ierr = MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5404abbbba34SStefano Zampini ierr = MatDestroy(&coarse_submat_dense);CHKERRQ(ierr); 5405abbbba34SStefano Zampini 54066e683305SStefano Zampini /* compute dofs splitting and neumann boundaries for coarse dofs */ 540730368db7SStefano Zampini if (multilevel_allowed && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local))) { /* protects from unneded computations */ 54086e683305SStefano Zampini PetscInt *tidxs,*tidxs2,nout,tsize,i; 54096e683305SStefano Zampini const PetscInt *idxs; 54106e683305SStefano Zampini ISLocalToGlobalMapping tmap; 54116e683305SStefano Zampini 54126e683305SStefano Zampini /* create map between primal indices (in local representative ordering) and local primal numbering */ 54130be93d54SStefano Zampini ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);CHKERRQ(ierr); 54146e683305SStefano Zampini /* allocate space for temporary storage */ 5415854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs);CHKERRQ(ierr); 5416854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->local_primal_size,&tidxs2);CHKERRQ(ierr); 54176e683305SStefano Zampini /* allocate for IS array */ 54186e683305SStefano Zampini nisdofs = pcbddc->n_ISForDofsLocal; 54196e683305SStefano Zampini nisneu = !!pcbddc->NeumannBoundariesLocal; 542030368db7SStefano Zampini nisvert = 0; 542130368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 542230368db7SStefano Zampini nisvert = 1; 542330368db7SStefano Zampini } 542430368db7SStefano Zampini nis = nisdofs + nisneu + nisvert; 5425854ce69bSBarry Smith ierr = PetscMalloc1(nis,&isarray);CHKERRQ(ierr); 54266e683305SStefano Zampini /* dofs splitting */ 54276e683305SStefano Zampini for (i=0;i<nisdofs;i++) { 54286e683305SStefano Zampini /* ierr = ISView(pcbddc->ISForDofsLocal[i],0);CHKERRQ(ierr); */ 54296e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);CHKERRQ(ierr); 54306e683305SStefano Zampini ierr = ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54316e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54326e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);CHKERRQ(ierr); 54336e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 543430368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);CHKERRQ(ierr); 54356e683305SStefano Zampini /* ierr = ISView(isarray[i],0);CHKERRQ(ierr); */ 54366e683305SStefano Zampini } 54376e683305SStefano Zampini /* neumann boundaries */ 54386e683305SStefano Zampini if (pcbddc->NeumannBoundariesLocal) { 54396e683305SStefano Zampini /* ierr = ISView(pcbddc->NeumannBoundariesLocal,0);CHKERRQ(ierr); */ 54406e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);CHKERRQ(ierr); 54416e683305SStefano Zampini ierr = ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54426e683305SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 54436e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);CHKERRQ(ierr); 54446e683305SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 544530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);CHKERRQ(ierr); 54466e683305SStefano Zampini /* ierr = ISView(isarray[nisdofs],0);CHKERRQ(ierr); */ 54476e683305SStefano Zampini } 544830368db7SStefano Zampini /* primal vertices (benign) */ 544930368db7SStefano Zampini if (pcbddc->benign_saddle_point && pcbddc->user_primal_vertices_local) { 545030368db7SStefano Zampini ierr = ISGetLocalSize(pcbddc->user_primal_vertices_local,&tsize);CHKERRQ(ierr); 545130368db7SStefano Zampini ierr = ISGetIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 545230368db7SStefano Zampini ierr = ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);CHKERRQ(ierr); 545330368db7SStefano Zampini ierr = ISRestoreIndices(pcbddc->user_primal_vertices_local,&idxs);CHKERRQ(ierr); 545430368db7SStefano Zampini ierr = ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);CHKERRQ(ierr); 545530368db7SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nis-1]);CHKERRQ(ierr); 545630368db7SStefano Zampini } 54576e683305SStefano Zampini /* free memory */ 54586e683305SStefano Zampini ierr = PetscFree(tidxs);CHKERRQ(ierr); 54596e683305SStefano Zampini ierr = PetscFree(tidxs2);CHKERRQ(ierr); 54606e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&tmap);CHKERRQ(ierr); 54616e683305SStefano Zampini } else { 54626e683305SStefano Zampini nis = 0; 54636e683305SStefano Zampini nisdofs = 0; 54646e683305SStefano Zampini nisneu = 0; 546530368db7SStefano Zampini nisvert = 0; 54666e683305SStefano Zampini isarray = NULL; 54676e683305SStefano Zampini } 54686e683305SStefano Zampini /* destroy no longer needed map */ 54696e683305SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&coarse_islg);CHKERRQ(ierr); 54706e683305SStefano Zampini 54716e683305SStefano Zampini /* restrict on coarse candidates (if needed) */ 54726e683305SStefano Zampini coarse_mat_is = NULL; 54736e683305SStefano Zampini if (csin) { 54746e683305SStefano Zampini if (!pcbddc->coarse_subassembling_init ) { /* creates subassembling init pattern if not present */ 547574e2c79eSStefano Zampini if (redist) { 547674e2c79eSStefano Zampini PetscMPIInt rank; 5477779c1cceSStefano Zampini PetscInt spc,n_spc_p1,dest[1],destsize; 547874e2c79eSStefano Zampini 547974e2c79eSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 548058da7f69SStefano Zampini spc = active_procs/ncoarse; 548158da7f69SStefano Zampini n_spc_p1 = active_procs%ncoarse; 5482779c1cceSStefano Zampini if (im_active) { 5483779c1cceSStefano Zampini destsize = 1; 548474e2c79eSStefano Zampini if (rank > n_spc_p1*(spc+1)-1) { 548574e2c79eSStefano Zampini dest[0] = n_spc_p1+(rank-(n_spc_p1*(spc+1)))/spc; 548674e2c79eSStefano Zampini } else { 548774e2c79eSStefano Zampini dest[0] = rank/(spc+1); 548874e2c79eSStefano Zampini } 548974e2c79eSStefano Zampini } else { 5490779c1cceSStefano Zampini destsize = 0; 54916e683305SStefano Zampini } 5492779c1cceSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),destsize,dest,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5493779c1cceSStefano Zampini } else if (csin_type_simple) { 54946e683305SStefano Zampini PetscMPIInt rank; 54956e683305SStefano Zampini PetscInt issize,isidx; 5496779c1cceSStefano Zampini 54976e683305SStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);CHKERRQ(ierr); 54986e683305SStefano Zampini if (im_active) { 54996e683305SStefano Zampini issize = 1; 55006e683305SStefano Zampini isidx = (PetscInt)rank; 55016e683305SStefano Zampini } else { 55026e683305SStefano Zampini issize = 0; 55036e683305SStefano Zampini isidx = -1; 55046e683305SStefano Zampini } 55056e683305SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),issize,&isidx,PETSC_COPY_VALUES,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 5506779c1cceSStefano Zampini } else { /* get a suitable subassembling pattern from MATIS code */ 5507b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(t_coarse_mat_is,ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling_init);CHKERRQ(ierr); 55086e683305SStefano Zampini } 5509779c1cceSStefano Zampini 5510779c1cceSStefano Zampini /* we need to shift on coarse candidates either if we are not redistributing or we are redistributing and we have enough void processes */ 5511779c1cceSStefano Zampini if (!redist || ncoarse <= void_procs) { 5512779c1cceSStefano Zampini PetscInt ncoarse_cand,tissize,*nisindices; 5513779c1cceSStefano Zampini PetscInt *coarse_candidates; 5514779c1cceSStefano Zampini const PetscInt* tisindices; 5515779c1cceSStefano Zampini 5516779c1cceSStefano Zampini /* get coarse candidates' ranks in pc communicator */ 5517779c1cceSStefano Zampini ierr = PetscMalloc1(all_procs,&coarse_candidates);CHKERRQ(ierr); 5518779c1cceSStefano Zampini ierr = MPI_Allgather(&im_active,1,MPIU_INT,coarse_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 5519779c1cceSStefano Zampini for (i=0,ncoarse_cand=0;i<all_procs;i++) { 5520779c1cceSStefano Zampini if (!coarse_candidates[i]) { 5521779c1cceSStefano Zampini coarse_candidates[ncoarse_cand++]=i; 5522779c1cceSStefano Zampini } 5523779c1cceSStefano Zampini } 5524779c1cceSStefano Zampini if (ncoarse_cand < ncoarse) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen! %d < %d",ncoarse_cand,ncoarse); 5525779c1cceSStefano Zampini 5526779c1cceSStefano Zampini 55276e683305SStefano Zampini if (pcbddc->dbg_flag) { 55286e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55296e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init (before shift)\n");CHKERRQ(ierr); 55306e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55316e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse candidates\n");CHKERRQ(ierr); 5532779c1cceSStefano Zampini for (i=0;i<ncoarse_cand;i++) { 55336e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"%d ",coarse_candidates[i]);CHKERRQ(ierr); 55346e683305SStefano Zampini } 55356e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"\n");CHKERRQ(ierr); 55366e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55376e683305SStefano Zampini } 55386e683305SStefano Zampini /* shift the pattern on coarse candidates */ 55396e683305SStefano Zampini ierr = ISGetLocalSize(pcbddc->coarse_subassembling_init,&tissize);CHKERRQ(ierr); 55406e683305SStefano Zampini ierr = ISGetIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 5541854ce69bSBarry Smith ierr = PetscMalloc1(tissize,&nisindices);CHKERRQ(ierr); 55426e683305SStefano Zampini for (i=0;i<tissize;i++) nisindices[i] = coarse_candidates[tisindices[i]]; 55436e683305SStefano Zampini ierr = ISRestoreIndices(pcbddc->coarse_subassembling_init,&tisindices);CHKERRQ(ierr); 55446e683305SStefano Zampini ierr = ISGeneralSetIndices(pcbddc->coarse_subassembling_init,tissize,nisindices,PETSC_OWN_POINTER);CHKERRQ(ierr); 55456e683305SStefano Zampini ierr = PetscFree(coarse_candidates);CHKERRQ(ierr); 55466e683305SStefano Zampini } 55476e683305SStefano Zampini if (pcbddc->dbg_flag) { 55486e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55496e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init\n");CHKERRQ(ierr); 55506e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling_init,pcbddc->dbg_viewer);CHKERRQ(ierr); 55516e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55526e683305SStefano Zampini } 5553779c1cceSStefano Zampini } 55546e683305SStefano Zampini /* get temporary coarse mat in IS format restricted on coarse procs (plus additional index sets of isarray) */ 555553a05cb3SStefano Zampini if (multilevel_allowed) { /* we need to keep tracking of void processes for future placements */ 555653a05cb3SStefano 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); 555753a05cb3SStefano Zampini } else { /* this is the last level, so use just receiving processes in subcomm */ 555853a05cb3SStefano 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); 555953a05cb3SStefano Zampini } 55606e683305SStefano Zampini } else { 55616e683305SStefano Zampini if (pcbddc->dbg_flag) { 55626e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 55636e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Subassembling pattern init not needed\n");CHKERRQ(ierr); 55646e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 55656e683305SStefano Zampini } 55666e683305SStefano Zampini ierr = PetscObjectReference((PetscObject)t_coarse_mat_is);CHKERRQ(ierr); 55676e683305SStefano Zampini coarse_mat_is = t_coarse_mat_is; 55686e683305SStefano Zampini } 55696e683305SStefano Zampini 55706e683305SStefano Zampini /* create local to global scatters for coarse problem */ 557168457ee5SStefano Zampini if (compute_vecs) { 55726e683305SStefano Zampini PetscInt lrows; 55736e683305SStefano Zampini ierr = VecDestroy(&pcbddc->coarse_vec);CHKERRQ(ierr); 55746e683305SStefano Zampini if (coarse_mat_is) { 55756e683305SStefano Zampini ierr = MatGetLocalSize(coarse_mat_is,&lrows,NULL);CHKERRQ(ierr); 55766e683305SStefano Zampini } else { 55776e683305SStefano Zampini lrows = 0; 55786e683305SStefano Zampini } 55796e683305SStefano Zampini ierr = VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);CHKERRQ(ierr); 55806e683305SStefano Zampini ierr = VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);CHKERRQ(ierr); 55816e683305SStefano Zampini ierr = VecSetType(pcbddc->coarse_vec,VECSTANDARD);CHKERRQ(ierr); 55826e683305SStefano Zampini ierr = VecScatterDestroy(&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55836e683305SStefano Zampini ierr = VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);CHKERRQ(ierr); 55846e683305SStefano Zampini } 55856e683305SStefano Zampini ierr = ISDestroy(&coarse_is);CHKERRQ(ierr); 55866e683305SStefano Zampini ierr = MatDestroy(&t_coarse_mat_is);CHKERRQ(ierr); 5587c8587f34SStefano Zampini 5588f9eb5b7dSStefano Zampini /* set defaults for coarse KSP and PC */ 5589f9eb5b7dSStefano Zampini if (multilevel_allowed) { 5590f9eb5b7dSStefano Zampini coarse_ksp_type = KSPRICHARDSON; 5591f9eb5b7dSStefano Zampini coarse_pc_type = PCBDDC; 5592f9eb5b7dSStefano Zampini } else { 5593f9eb5b7dSStefano Zampini coarse_ksp_type = KSPPREONLY; 5594f9eb5b7dSStefano Zampini coarse_pc_type = PCREDUNDANT; 5595c8587f34SStefano Zampini } 5596c8587f34SStefano Zampini 55976e683305SStefano Zampini /* print some info if requested */ 55986e683305SStefano Zampini if (pcbddc->dbg_flag) { 55996e683305SStefano Zampini if (!multilevel_allowed) { 56006e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 56016e683305SStefano Zampini if (multilevel_requested) { 56026e683305SStefano 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); 56036e683305SStefano Zampini } else if (pcbddc->max_levels) { 56046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%d)\n",pcbddc->max_levels);CHKERRQ(ierr); 56056e683305SStefano Zampini } 56066e683305SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 56076e683305SStefano Zampini } 56086e683305SStefano Zampini } 56096e683305SStefano Zampini 5610f9eb5b7dSStefano Zampini /* create the coarse KSP object only once with defaults */ 56116e683305SStefano Zampini if (coarse_mat_is) { 56126e683305SStefano Zampini MatReuse coarse_mat_reuse; 56136a1308c2SStefano Zampini PetscViewer dbg_viewer = NULL; 56146e683305SStefano Zampini if (pcbddc->dbg_flag) { 56156e683305SStefano Zampini dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat_is)); 56166e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 56176e683305SStefano Zampini } 5618f9eb5b7dSStefano Zampini if (!pcbddc->coarse_ksp) { 5619312be037SStefano Zampini char prefix[256],str_level[16]; 5620e604994aSStefano Zampini size_t len; 56216e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)coarse_mat_is),&pcbddc->coarse_ksp);CHKERRQ(ierr); 5622422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 5623c8587f34SStefano Zampini ierr = PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);CHKERRQ(ierr); 5624f9eb5b7dSStefano Zampini ierr = KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);CHKERRQ(ierr); 56255f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat_is,coarse_mat_is);CHKERRQ(ierr); 5626c8587f34SStefano Zampini ierr = KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);CHKERRQ(ierr); 56276e683305SStefano Zampini ierr = KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);CHKERRQ(ierr); 5628c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 5629c8587f34SStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5630e604994aSStefano Zampini /* prefix */ 5631e604994aSStefano Zampini ierr = PetscStrcpy(prefix,"");CHKERRQ(ierr); 5632e604994aSStefano Zampini ierr = PetscStrcpy(str_level,"");CHKERRQ(ierr); 5633e604994aSStefano Zampini if (!pcbddc->current_level) { 5634e604994aSStefano Zampini ierr = PetscStrcpy(prefix,((PetscObject)pc)->prefix);CHKERRQ(ierr); 5635e604994aSStefano Zampini ierr = PetscStrcat(prefix,"pc_bddc_coarse_");CHKERRQ(ierr); 5636c8587f34SStefano Zampini } else { 5637e604994aSStefano Zampini ierr = PetscStrlen(((PetscObject)pc)->prefix,&len);CHKERRQ(ierr); 5638312be037SStefano Zampini if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */ 5639312be037SStefano Zampini if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */ 564034d6797cSStefano Zampini ierr = PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);CHKERRQ(ierr); 5641312be037SStefano Zampini sprintf(str_level,"l%d_",(int)(pcbddc->current_level)); 5642e604994aSStefano Zampini ierr = PetscStrcat(prefix,str_level);CHKERRQ(ierr); 5643e604994aSStefano Zampini } 5644e604994aSStefano Zampini ierr = KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);CHKERRQ(ierr); 56453e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 56463e3c6dadSStefano Zampini ierr = PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);CHKERRQ(ierr); 56473e3c6dadSStefano Zampini ierr = PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);CHKERRQ(ierr); 56483e3c6dadSStefano Zampini ierr = PCBDDCSetLevels(pc_temp,pcbddc->max_levels);CHKERRQ(ierr); 5649f9eb5b7dSStefano Zampini /* allow user customization */ 5650f9eb5b7dSStefano Zampini ierr = KSPSetFromOptions(pcbddc->coarse_ksp);CHKERRQ(ierr); 56513e3c6dadSStefano Zampini } 56523e3c6dadSStefano Zampini /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */ 565351bea450SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&pc_temp);CHKERRQ(ierr); 56543e3c6dadSStefano Zampini if (nisdofs) { 56553e3c6dadSStefano Zampini ierr = PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);CHKERRQ(ierr); 56563e3c6dadSStefano Zampini for (i=0;i<nisdofs;i++) { 56573e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[i]);CHKERRQ(ierr); 56583e3c6dadSStefano Zampini } 56593e3c6dadSStefano Zampini } 56603e3c6dadSStefano Zampini if (nisneu) { 56613e3c6dadSStefano Zampini ierr = PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);CHKERRQ(ierr); 56623e3c6dadSStefano Zampini ierr = ISDestroy(&isarray[nisdofs]);CHKERRQ(ierr); 5663312be037SStefano Zampini } 566430368db7SStefano Zampini if (nisvert) { 566530368db7SStefano Zampini ierr = PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);CHKERRQ(ierr); 566630368db7SStefano Zampini ierr = ISDestroy(&isarray[nis-1]);CHKERRQ(ierr); 566730368db7SStefano Zampini } 5668f9eb5b7dSStefano Zampini 5669f9eb5b7dSStefano Zampini /* get some info after set from options */ 5670f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCNN,&isnn);CHKERRQ(ierr); 5671f9eb5b7dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);CHKERRQ(ierr); 56724f3a063dSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);CHKERRQ(ierr); 56736e683305SStefano Zampini if (isbddc && !multilevel_allowed) { /* multilevel can only be requested via pc_bddc_set_levels */ 5674f9eb5b7dSStefano Zampini ierr = PCSetType(pc_temp,coarse_pc_type);CHKERRQ(ierr); 5675f9eb5b7dSStefano Zampini isbddc = PETSC_FALSE; 5676f9eb5b7dSStefano Zampini } 567739f0f02cSStefano Zampini ierr = PCFactorSetReuseFill(pc_temp,PETSC_TRUE);CHKERRQ(ierr); 56784f3a063dSStefano Zampini if (isredundant) { 56794f3a063dSStefano Zampini KSP inner_ksp; 56804f3a063dSStefano Zampini PC inner_pc; 56814f3a063dSStefano Zampini ierr = PCRedundantGetKSP(pc_temp,&inner_ksp);CHKERRQ(ierr); 56824f3a063dSStefano Zampini ierr = KSPGetPC(inner_ksp,&inner_pc);CHKERRQ(ierr); 56834f3a063dSStefano Zampini ierr = PCFactorSetReuseFill(inner_pc,PETSC_TRUE);CHKERRQ(ierr); 56844f3a063dSStefano Zampini } 5685f9eb5b7dSStefano Zampini 5686f9eb5b7dSStefano Zampini /* assemble coarse matrix */ 5687fa7f1dd8SStefano Zampini if (coarse_reuse) { 568881d9aea3SBarry Smith ierr = KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);CHKERRQ(ierr); 5689fa7f1dd8SStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat);CHKERRQ(ierr); 56906e683305SStefano Zampini coarse_mat_reuse = MAT_REUSE_MATRIX; 5691fa7f1dd8SStefano Zampini } else { 56926e683305SStefano Zampini coarse_mat_reuse = MAT_INITIAL_MATRIX; 5693fa7f1dd8SStefano Zampini } 5694c8587f34SStefano Zampini if (isbddc || isnn) { 5695d4d8cf7bSStefano Zampini if (isbddc) { /* currently there's no API for this */ 5696d4d8cf7bSStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc_temp->data; 5697d4d8cf7bSStefano Zampini pcbddc->detect_disconnected = PETSC_TRUE; 5698d4d8cf7bSStefano Zampini } 569922bc73bbSStefano Zampini if (pcbddc->coarsening_ratio > 1) { 570070cf5478SStefano Zampini if (!pcbddc->coarse_subassembling) { /* subassembling info is not present */ 5701b0c7d250SStefano Zampini ierr = MatISGetSubassemblingPattern(coarse_mat_is,active_procs/pcbddc->coarsening_ratio,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling);CHKERRQ(ierr); 570222b6e8a2SStefano Zampini if (pcbddc->dbg_flag) { 57036e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 57046e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Subassembling pattern\n");CHKERRQ(ierr); 57056e683305SStefano Zampini ierr = ISView(pcbddc->coarse_subassembling,dbg_viewer);CHKERRQ(ierr); 57066e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 570722b6e8a2SStefano Zampini } 570870cf5478SStefano Zampini } 570953a05cb3SStefano Zampini ierr = MatISSubassemble(coarse_mat_is,pcbddc->coarse_subassembling,0,PETSC_FALSE,PETSC_FALSE,coarse_mat_reuse,&coarse_mat,0,NULL);CHKERRQ(ierr); 571070cf5478SStefano Zampini } else { 571122bc73bbSStefano Zampini ierr = PetscObjectReference((PetscObject)coarse_mat_is);CHKERRQ(ierr); 571222bc73bbSStefano Zampini coarse_mat = coarse_mat_is; 571322bc73bbSStefano Zampini } 571422bc73bbSStefano Zampini } else { 57152e1e5fa4SStefano Zampini ierr = MatISGetMPIXAIJ(coarse_mat_is,coarse_mat_reuse,&coarse_mat);CHKERRQ(ierr); 5716c8587f34SStefano Zampini } 5717c8587f34SStefano Zampini ierr = MatDestroy(&coarse_mat_is);CHKERRQ(ierr); 5718c8587f34SStefano Zampini 57193301b35fSStefano Zampini /* propagate symmetry info of coarse matrix */ 57205a16e3a0SStefano Zampini ierr = MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 57213301b35fSStefano Zampini if (pc->pmat->symmetric_set) { 57223301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);CHKERRQ(ierr); 57233301b35fSStefano Zampini } 57243301b35fSStefano Zampini if (pc->pmat->hermitian_set) { 57253301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);CHKERRQ(ierr); 57263301b35fSStefano Zampini } 57273301b35fSStefano Zampini if (pc->pmat->spd_set) { 57283301b35fSStefano Zampini ierr = MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);CHKERRQ(ierr); 57293301b35fSStefano Zampini } 57306e683305SStefano Zampini /* set operators */ 57315f76c7aeSStefano Zampini ierr = KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 57326e683305SStefano Zampini if (pcbddc->dbg_flag) { 57336e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);CHKERRQ(ierr); 57346e683305SStefano Zampini } 57356e683305SStefano Zampini } else { /* processes non partecipating to coarse solver (if any) */ 57366e683305SStefano Zampini coarse_mat = 0; 57376e683305SStefano Zampini } 57386e683305SStefano Zampini ierr = PetscFree(isarray);CHKERRQ(ierr); 5739b1ecc7b1SStefano Zampini #if 0 5740b9b85e73SStefano Zampini { 5741b9b85e73SStefano Zampini PetscViewer viewer; 5742b9b85e73SStefano Zampini char filename[256]; 5743b1ecc7b1SStefano Zampini sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level); 5744b1ecc7b1SStefano Zampini ierr = PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);CHKERRQ(ierr); 5745b9b85e73SStefano Zampini ierr = PetscViewerSetFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 5746b9b85e73SStefano Zampini ierr = MatView(coarse_mat,viewer);CHKERRQ(ierr); 5747b9b85e73SStefano Zampini ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); 5748b9b85e73SStefano Zampini } 5749b9b85e73SStefano Zampini #endif 5750c8587f34SStefano Zampini 5751c8587f34SStefano Zampini /* Compute coarse null space (special handling by BDDC only) */ 5752298c0119SStefano Zampini #if 0 5753c8587f34SStefano Zampini if (pcbddc->NullSpace) { 5754c8587f34SStefano Zampini ierr = PCBDDCNullSpaceAssembleCoarse(pc,coarse_mat,&CoarseNullSpace);CHKERRQ(ierr); 575598a51de6SStefano Zampini } 5756298c0119SStefano Zampini #endif 5757b0f5fe93SStefano Zampini /* hack */ 575898a51de6SStefano Zampini if (pcbddc->coarse_ksp) { 575998a51de6SStefano Zampini Vec crhs,csol; 576004708bb6SStefano Zampini 5761f347579bSStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&csol);CHKERRQ(ierr); 5762f347579bSStefano Zampini ierr = KSPGetRhs(pcbddc->coarse_ksp,&crhs);CHKERRQ(ierr); 5763f347579bSStefano Zampini if (!csol) { 57642a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);CHKERRQ(ierr); 5765f9eb5b7dSStefano Zampini } 5766f347579bSStefano Zampini if (!crhs) { 57672a7a6963SBarry Smith ierr = MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));CHKERRQ(ierr); 5768f347579bSStefano Zampini } 5769b0f5fe93SStefano Zampini } 5770b0f5fe93SStefano Zampini 5771b0f5fe93SStefano Zampini /* compute null space for coarse solver if the benign trick has been requested */ 5772b0f5fe93SStefano Zampini if (pcbddc->benign_null) { 5773b0f5fe93SStefano Zampini 5774b0f5fe93SStefano Zampini ierr = VecSet(pcbddc->vec1_P,0.);CHKERRQ(ierr); 57754f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 57764f1b2e48SStefano Zampini ierr = VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);CHKERRQ(ierr); 57774f1b2e48SStefano Zampini } 5778b0f5fe93SStefano Zampini ierr = VecAssemblyBegin(pcbddc->vec1_P);CHKERRQ(ierr); 5779b0f5fe93SStefano Zampini ierr = VecAssemblyEnd(pcbddc->vec1_P);CHKERRQ(ierr); 5780b0f5fe93SStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5781b0f5fe93SStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5782b0f5fe93SStefano Zampini if (coarse_mat) { 5783b0f5fe93SStefano Zampini Vec nullv; 5784b0f5fe93SStefano Zampini PetscScalar *array,*array2; 5785b0f5fe93SStefano Zampini PetscInt nl; 5786b0f5fe93SStefano Zampini 5787b0f5fe93SStefano Zampini ierr = MatCreateVecs(coarse_mat,&nullv,NULL);CHKERRQ(ierr); 5788b0f5fe93SStefano Zampini ierr = VecGetLocalSize(nullv,&nl);CHKERRQ(ierr); 5789b0f5fe93SStefano Zampini ierr = VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5790b0f5fe93SStefano Zampini ierr = VecGetArray(nullv,&array2);CHKERRQ(ierr); 5791b0f5fe93SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(*array));CHKERRQ(ierr); 5792b0f5fe93SStefano Zampini ierr = VecRestoreArray(nullv,&array2);CHKERRQ(ierr); 5793b0f5fe93SStefano Zampini ierr = VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);CHKERRQ(ierr); 5794b0f5fe93SStefano Zampini ierr = VecNormalize(nullv,NULL);CHKERRQ(ierr); 5795b0f5fe93SStefano Zampini ierr = MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);CHKERRQ(ierr); 5796b0f5fe93SStefano Zampini ierr = VecDestroy(&nullv);CHKERRQ(ierr); 5797b0f5fe93SStefano Zampini } 5798b0f5fe93SStefano Zampini } 5799b0f5fe93SStefano Zampini 5800b0f5fe93SStefano Zampini if (pcbddc->coarse_ksp) { 5801b0f5fe93SStefano Zampini PetscBool ispreonly; 5802b0f5fe93SStefano Zampini 5803b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5804b0f5fe93SStefano Zampini PetscBool isnull; 5805b0f5fe93SStefano Zampini ierr = MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);CHKERRQ(ierr); 5806d4d8cf7bSStefano Zampini if (0) { 580730368db7SStefano Zampini if (isbddc && !pcbddc->benign_saddle_point) { 5808b0f5fe93SStefano Zampini ierr = PCBDDCSetNullSpace(pc_temp,CoarseNullSpace);CHKERRQ(ierr); 5809b0f5fe93SStefano Zampini } else { 5810b0f5fe93SStefano Zampini ierr = MatSetNullSpace(coarse_mat,CoarseNullSpace);CHKERRQ(ierr); 5811b0f5fe93SStefano Zampini } 5812b0f5fe93SStefano Zampini } else { 5813b0f5fe93SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5814b0f5fe93SStefano Zampini } 5815b0f5fe93SStefano Zampini } 5816b0f5fe93SStefano Zampini /* setup coarse ksp */ 5817b0f5fe93SStefano Zampini ierr = KSPSetUp(pcbddc->coarse_ksp);CHKERRQ(ierr); 5818cbcc2c2aSStefano Zampini /* Check coarse problem if in debug mode or if solving with an iterative method */ 5819cbcc2c2aSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);CHKERRQ(ierr); 58206e683305SStefano Zampini if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) { 5821c8587f34SStefano Zampini KSP check_ksp; 58222b510759SStefano Zampini KSPType check_ksp_type; 5823c8587f34SStefano Zampini PC check_pc; 58246e683305SStefano Zampini Vec check_vec,coarse_vec; 58256a1308c2SStefano Zampini PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0; 58262b510759SStefano Zampini PetscInt its; 58276e683305SStefano Zampini PetscBool compute_eigs; 58286e683305SStefano Zampini PetscReal *eigs_r,*eigs_c; 58296e683305SStefano Zampini PetscInt neigs; 58308e185a42SStefano Zampini const char *prefix; 5831c8587f34SStefano Zampini 58322b510759SStefano Zampini /* Create ksp object suitable for estimation of extreme eigenvalues */ 58336e683305SStefano Zampini ierr = KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);CHKERRQ(ierr); 5834422a814eSBarry Smith ierr = KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);CHKERRQ(ierr); 583523ee1639SBarry Smith ierr = KSPSetOperators(check_ksp,coarse_mat,coarse_mat);CHKERRQ(ierr); 5836f4ddd8eeSStefano Zampini ierr = KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);CHKERRQ(ierr); 58372b510759SStefano Zampini if (ispreonly) { 58382b510759SStefano Zampini check_ksp_type = KSPPREONLY; 58396e683305SStefano Zampini compute_eigs = PETSC_FALSE; 58402b510759SStefano Zampini } else { 5841cbcc2c2aSStefano Zampini check_ksp_type = KSPGMRES; 58426e683305SStefano Zampini compute_eigs = PETSC_TRUE; 5843c8587f34SStefano Zampini } 5844c8587f34SStefano Zampini ierr = KSPSetType(check_ksp,check_ksp_type);CHKERRQ(ierr); 58456e683305SStefano Zampini ierr = KSPSetComputeSingularValues(check_ksp,compute_eigs);CHKERRQ(ierr); 58466e683305SStefano Zampini ierr = KSPSetComputeEigenvalues(check_ksp,compute_eigs);CHKERRQ(ierr); 58476e683305SStefano Zampini ierr = KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);CHKERRQ(ierr); 5848a7dc3881SStefano Zampini ierr = KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);CHKERRQ(ierr); 5849a7dc3881SStefano Zampini ierr = KSPSetOptionsPrefix(check_ksp,prefix);CHKERRQ(ierr); 5850a7dc3881SStefano Zampini ierr = KSPAppendOptionsPrefix(check_ksp,"check_");CHKERRQ(ierr); 5851a7dc3881SStefano Zampini ierr = KSPSetFromOptions(check_ksp);CHKERRQ(ierr); 5852c8587f34SStefano Zampini ierr = KSPSetUp(check_ksp);CHKERRQ(ierr); 5853c8587f34SStefano Zampini ierr = KSPGetPC(pcbddc->coarse_ksp,&check_pc);CHKERRQ(ierr); 5854c8587f34SStefano Zampini ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr); 5855c8587f34SStefano Zampini /* create random vec */ 58566e683305SStefano Zampini ierr = KSPGetSolution(pcbddc->coarse_ksp,&coarse_vec);CHKERRQ(ierr); 58576e683305SStefano Zampini ierr = VecDuplicate(coarse_vec,&check_vec);CHKERRQ(ierr); 5858c8587f34SStefano Zampini ierr = VecSetRandom(check_vec,NULL);CHKERRQ(ierr); 5859c8587f34SStefano Zampini if (CoarseNullSpace) { 5860c8587f34SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,check_vec);CHKERRQ(ierr); 5861c8587f34SStefano Zampini } 58626e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 5863c8587f34SStefano Zampini /* solve coarse problem */ 58646e683305SStefano Zampini ierr = KSPSolve(check_ksp,coarse_vec,coarse_vec);CHKERRQ(ierr); 5865c8587f34SStefano Zampini if (CoarseNullSpace) { 58666e683305SStefano Zampini ierr = MatNullSpaceRemove(CoarseNullSpace,coarse_vec);CHKERRQ(ierr); 5867c8587f34SStefano Zampini } 5868cbcc2c2aSStefano Zampini /* set eigenvalue estimation if preonly has not been requested */ 58696e683305SStefano Zampini if (compute_eigs) { 5870854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);CHKERRQ(ierr); 5871854ce69bSBarry Smith ierr = PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);CHKERRQ(ierr); 58726e683305SStefano Zampini ierr = KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);CHKERRQ(ierr); 58736e683305SStefano Zampini lambda_max = eigs_r[neigs-1]; 58746e683305SStefano Zampini lambda_min = eigs_r[0]; 58756e683305SStefano Zampini if (pcbddc->use_coarse_estimates) { 58766e683305SStefano Zampini if (lambda_max>lambda_min) { 5877cbcc2c2aSStefano Zampini ierr = KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max,lambda_min);CHKERRQ(ierr); 5878cbcc2c2aSStefano Zampini ierr = KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));CHKERRQ(ierr); 5879cbcc2c2aSStefano Zampini } 5880c8587f34SStefano Zampini } 5881c8587f34SStefano Zampini } 5882cbcc2c2aSStefano Zampini 5883c8587f34SStefano Zampini /* check coarse problem residual error */ 58846e683305SStefano Zampini if (pcbddc->dbg_flag) { 58856e683305SStefano Zampini PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp)); 58866e683305SStefano Zampini ierr = PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 58876e683305SStefano Zampini ierr = VecAXPY(check_vec,-1.0,coarse_vec);CHKERRQ(ierr); 5888c8587f34SStefano Zampini ierr = VecNorm(check_vec,NORM_INFINITY,&infty_error);CHKERRQ(ierr); 58896e683305SStefano Zampini ierr = MatMult(coarse_mat,check_vec,coarse_vec);CHKERRQ(ierr); 58906e683305SStefano Zampini ierr = VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);CHKERRQ(ierr); 5891c8587f34SStefano Zampini ierr = VecDestroy(&check_vec);CHKERRQ(ierr); 5892779c1cceSStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);CHKERRQ(ierr); 58936e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);CHKERRQ(ierr); 58946e683305SStefano Zampini ierr = PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);CHKERRQ(ierr); 58956e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);CHKERRQ(ierr); 58966e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);CHKERRQ(ierr); 5897b0f5fe93SStefano Zampini if (CoarseNullSpace) { 5898b0f5fe93SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");CHKERRQ(ierr); 5899b0f5fe93SStefano Zampini } 59006e683305SStefano Zampini if (compute_eigs) { 59016e683305SStefano Zampini PetscReal lambda_max_s,lambda_min_s; 5902deec49d1SStefano Zampini ierr = KSPGetType(check_ksp,&check_ksp_type);CHKERRQ(ierr); 5903c8587f34SStefano Zampini ierr = KSPGetIterationNumber(check_ksp,&its);CHKERRQ(ierr); 59046e683305SStefano Zampini ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);CHKERRQ(ierr); 59056e683305SStefano 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); 59066e683305SStefano Zampini for (i=0;i<neigs;i++) { 59076e683305SStefano Zampini ierr = PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);CHKERRQ(ierr); 5908c8587f34SStefano Zampini } 59096e683305SStefano Zampini } 59106e683305SStefano Zampini ierr = PetscViewerFlush(dbg_viewer);CHKERRQ(ierr); 59116e683305SStefano Zampini ierr = PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));CHKERRQ(ierr); 59126e683305SStefano Zampini } 5913c8587f34SStefano Zampini ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr); 59146e683305SStefano Zampini if (compute_eigs) { 59156e683305SStefano Zampini ierr = PetscFree(eigs_r);CHKERRQ(ierr); 59166e683305SStefano Zampini ierr = PetscFree(eigs_c);CHKERRQ(ierr); 5917c8587f34SStefano Zampini } 59186e683305SStefano Zampini } 59196e683305SStefano Zampini } 5920cbcc2c2aSStefano Zampini /* print additional info */ 5921cbcc2c2aSStefano Zampini if (pcbddc->dbg_flag) { 59226e683305SStefano Zampini /* waits until all processes reaches this point */ 59236e683305SStefano Zampini ierr = PetscBarrier((PetscObject)pc);CHKERRQ(ierr); 5924cbcc2c2aSStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %d\n",pcbddc->current_level);CHKERRQ(ierr); 5925cbcc2c2aSStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5926cbcc2c2aSStefano Zampini } 5927cbcc2c2aSStefano Zampini 59282b510759SStefano Zampini /* free memory */ 5929c8587f34SStefano Zampini ierr = MatNullSpaceDestroy(&CoarseNullSpace);CHKERRQ(ierr); 5930fdc635d7SStefano Zampini ierr = MatDestroy(&coarse_mat);CHKERRQ(ierr); 5931c8587f34SStefano Zampini PetscFunctionReturn(0); 5932c8587f34SStefano Zampini } 5933674ae819SStefano Zampini 5934f34684f1SStefano Zampini #undef __FUNCT__ 5935f34684f1SStefano Zampini #define __FUNCT__ "PCBDDCComputePrimalNumbering" 5936f34684f1SStefano Zampini PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n) 5937f34684f1SStefano Zampini { 5938f34684f1SStefano Zampini PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 5939f34684f1SStefano Zampini PC_IS* pcis = (PC_IS*)pc->data; 5940f34684f1SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 5941dc456d91SStefano Zampini IS subset,subset_mult,subset_n; 5942dc456d91SStefano Zampini PetscInt local_size,coarse_size=0; 594373be2a3aSStefano Zampini PetscInt *local_primal_indices=NULL; 5944dc456d91SStefano Zampini const PetscInt *t_local_primal_indices; 5945f34684f1SStefano Zampini PetscErrorCode ierr; 5946f34684f1SStefano Zampini 5947f34684f1SStefano Zampini PetscFunctionBegin; 5948f34684f1SStefano Zampini /* Compute global number of coarse dofs */ 59490e6343abSStefano Zampini if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) { 59500e6343abSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first"); 5951727cdba6SStefano Zampini } 5952dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);CHKERRQ(ierr); 59533bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);CHKERRQ(ierr); 5954dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5955dc456d91SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);CHKERRQ(ierr); 5956dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(subset,subset_mult,&coarse_size,&subset_n);CHKERRQ(ierr); 5957dc456d91SStefano Zampini ierr = ISDestroy(&subset);CHKERRQ(ierr); 5958dc456d91SStefano Zampini ierr = ISDestroy(&subset_mult);CHKERRQ(ierr); 5959dc456d91SStefano Zampini ierr = ISGetLocalSize(subset_n,&local_size);CHKERRQ(ierr); 59600e6343abSStefano Zampini if (local_size != pcbddc->local_primal_size) { 59610e6343abSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %d != %d",local_size,pcbddc->local_primal_size); 59620e6343abSStefano Zampini } 5963dc456d91SStefano Zampini ierr = PetscMalloc1(local_size,&local_primal_indices);CHKERRQ(ierr); 5964dc456d91SStefano Zampini ierr = ISGetIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5965dc456d91SStefano Zampini ierr = PetscMemcpy(local_primal_indices,t_local_primal_indices,local_size*sizeof(PetscInt));CHKERRQ(ierr); 5966dc456d91SStefano Zampini ierr = ISRestoreIndices(subset_n,&t_local_primal_indices);CHKERRQ(ierr); 5967dc456d91SStefano Zampini ierr = ISDestroy(&subset_n);CHKERRQ(ierr); 5968f34684f1SStefano Zampini 5969f34684f1SStefano Zampini /* check numbering */ 5970f34684f1SStefano Zampini if (pcbddc->dbg_flag) { 5971019a44ceSStefano Zampini PetscScalar coarsesum,*array,*array2; 5972dc456d91SStefano Zampini PetscInt i; 5973b9b85e73SStefano Zampini PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE; 5974f34684f1SStefano Zampini 5975f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 5976f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");CHKERRQ(ierr); 5977f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");CHKERRQ(ierr); 59781575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 5979019a44ceSStefano Zampini /* counter */ 5980019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5981019a44ceSStefano Zampini ierr = VecSet(pcis->vec1_N,1.0);CHKERRQ(ierr); 5982019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5983019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5984019a44ceSStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5985019a44ceSStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5986f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_N,0.0);CHKERRQ(ierr); 5987f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 5988727cdba6SStefano Zampini ierr = VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);CHKERRQ(ierr); 5989f34684f1SStefano Zampini } 5990f34684f1SStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 5991f34684f1SStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 5992f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 5993e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5994e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 5995e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5996e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 5997f34684f1SStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 5998019a44ceSStefano Zampini ierr = VecGetArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 5999f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6000019a44ceSStefano Zampini if (array[i] != 0.0 && array[i] != array2[i]) { 600175c01103SStefano Zampini PetscInt owned = (PetscInt)PetscRealPart(array[i]); 600275c01103SStefano Zampini PetscInt neigh = (PetscInt)PetscRealPart(array2[i]); 6003b9b85e73SStefano Zampini set_error = PETSC_TRUE; 6004019a44ceSStefano 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); 6005f34684f1SStefano Zampini } 6006f34684f1SStefano Zampini } 6007019a44ceSStefano Zampini ierr = VecRestoreArray(pcis->vec2_N,&array2);CHKERRQ(ierr); 6008b9b85e73SStefano Zampini ierr = MPI_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 6009f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6010f34684f1SStefano Zampini for (i=0;i<pcis->n;i++) { 6011f34684f1SStefano Zampini if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]); 6012f34684f1SStefano Zampini } 6013f34684f1SStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6014f34684f1SStefano Zampini ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 6015e176bc59SStefano Zampini ierr = VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6016e176bc59SStefano Zampini ierr = VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6017f34684f1SStefano Zampini ierr = VecSum(pcis->vec1_global,&coarsesum);CHKERRQ(ierr); 6018f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %d (%lf)\n",coarse_size,PetscRealPart(coarsesum));CHKERRQ(ierr); 6019b9b85e73SStefano Zampini if (pcbddc->dbg_flag > 1 || set_error_reduced) { 6020ca8b9ea9SStefano Zampini PetscInt *gidxs; 6021ca8b9ea9SStefano Zampini 6022ca8b9ea9SStefano Zampini ierr = PetscMalloc1(pcbddc->local_primal_size,&gidxs);CHKERRQ(ierr); 60233bbff08aSStefano Zampini ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);CHKERRQ(ierr); 6024f34684f1SStefano Zampini ierr = PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");CHKERRQ(ierr); 6025f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6026f34684f1SStefano Zampini ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);CHKERRQ(ierr); 6027f34684f1SStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) { 60284bc2dc4bSStefano 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); 6029f34684f1SStefano Zampini } 6030f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 6031ca8b9ea9SStefano Zampini ierr = PetscFree(gidxs);CHKERRQ(ierr); 6032f34684f1SStefano Zampini } 6033f34684f1SStefano Zampini ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr); 60341575c14dSBarry Smith ierr = PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);CHKERRQ(ierr); 6035302440fdSBarry Smith if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed"); 6036f34684f1SStefano Zampini } 60378bec7fa6SStefano Zampini /* ierr = PetscPrintf(PetscObjectComm((PetscObject)pc),"Size of coarse problem is %d\n",coarse_size);CHKERRQ(ierr); */ 6038f34684f1SStefano Zampini /* get back data */ 6039f34684f1SStefano Zampini *coarse_size_n = coarse_size; 6040f34684f1SStefano Zampini *local_primal_indices_n = local_primal_indices; 6041674ae819SStefano Zampini PetscFunctionReturn(0); 6042674ae819SStefano Zampini } 6043674ae819SStefano Zampini 6044e456f2a8SStefano Zampini #undef __FUNCT__ 6045e456f2a8SStefano Zampini #define __FUNCT__ "PCBDDCGlobalToLocal" 6046a7dc3881SStefano Zampini PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis) 6047e456f2a8SStefano Zampini { 6048e456f2a8SStefano Zampini IS localis_t; 6049a7dc3881SStefano Zampini PetscInt i,lsize,*idxs,n; 6050e456f2a8SStefano Zampini PetscScalar *vals; 6051e456f2a8SStefano Zampini PetscErrorCode ierr; 6052e456f2a8SStefano Zampini 6053e456f2a8SStefano Zampini PetscFunctionBegin; 6054a7dc3881SStefano Zampini /* get indices in local ordering exploiting local to global map */ 6055e456f2a8SStefano Zampini ierr = ISGetLocalSize(globalis,&lsize);CHKERRQ(ierr); 6056854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&vals);CHKERRQ(ierr); 6057e456f2a8SStefano Zampini for (i=0;i<lsize;i++) vals[i] = 1.0; 6058e456f2a8SStefano Zampini ierr = ISGetIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6059a7dc3881SStefano Zampini ierr = VecSet(gwork,0.0);CHKERRQ(ierr); 6060a7dc3881SStefano Zampini ierr = VecSet(lwork,0.0);CHKERRQ(ierr); 60611035eff8SStefano Zampini if (idxs) { /* multilevel guard */ 6062a7dc3881SStefano Zampini ierr = VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);CHKERRQ(ierr); 60631035eff8SStefano Zampini } 6064a7dc3881SStefano Zampini ierr = VecAssemblyBegin(gwork);CHKERRQ(ierr); 6065e456f2a8SStefano Zampini ierr = ISRestoreIndices(globalis,(const PetscInt**)&idxs);CHKERRQ(ierr); 6066e456f2a8SStefano Zampini ierr = PetscFree(vals);CHKERRQ(ierr); 6067a7dc3881SStefano Zampini ierr = VecAssemblyEnd(gwork);CHKERRQ(ierr); 6068a7dc3881SStefano Zampini /* now compute set in local ordering */ 6069a7dc3881SStefano Zampini ierr = VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6070a7dc3881SStefano Zampini ierr = VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6071a7dc3881SStefano Zampini ierr = VecGetArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6072a7dc3881SStefano Zampini ierr = VecGetSize(lwork,&n);CHKERRQ(ierr); 6073a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6074ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6075e456f2a8SStefano Zampini lsize++; 6076e456f2a8SStefano Zampini } 6077e456f2a8SStefano Zampini } 6078854ce69bSBarry Smith ierr = PetscMalloc1(lsize,&idxs);CHKERRQ(ierr); 6079a7dc3881SStefano Zampini for (i=0,lsize=0;i<n;i++) { 6080ff92baa0SMatthew G. Knepley if (PetscRealPart(vals[i]) > 0.5) { 6081e456f2a8SStefano Zampini idxs[lsize++] = i; 6082e456f2a8SStefano Zampini } 6083e456f2a8SStefano Zampini } 6084a7dc3881SStefano Zampini ierr = VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);CHKERRQ(ierr); 6085a7dc3881SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);CHKERRQ(ierr); 6086e456f2a8SStefano Zampini *localis = localis_t; 6087e456f2a8SStefano Zampini PetscFunctionReturn(0); 6088e456f2a8SStefano Zampini } 6089906d46d4SStefano Zampini 6090906d46d4SStefano Zampini /* the next two functions will be called in KSPMatMult if a change of basis has been requested */ 6091906d46d4SStefano Zampini #undef __FUNCT__ 6092906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMult_Private" 6093906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMult_Private(Mat A, Vec x, Vec y) 6094906d46d4SStefano Zampini { 6095906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6096906d46d4SStefano Zampini PetscErrorCode ierr; 6097906d46d4SStefano Zampini 6098906d46d4SStefano Zampini PetscFunctionBegin; 6099906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6100906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6101906d46d4SStefano Zampini ierr = MatMult(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6102906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6103906d46d4SStefano Zampini PetscFunctionReturn(0); 6104906d46d4SStefano Zampini } 6105906d46d4SStefano Zampini 6106906d46d4SStefano Zampini #undef __FUNCT__ 6107906d46d4SStefano Zampini #define __FUNCT__ "PCBDDCMatMultTranspose_Private" 6108906d46d4SStefano Zampini static PetscErrorCode PCBDDCMatMultTranspose_Private(Mat A, Vec x, Vec y) 6109906d46d4SStefano Zampini { 6110906d46d4SStefano Zampini PCBDDCChange_ctx change_ctx; 6111906d46d4SStefano Zampini PetscErrorCode ierr; 6112906d46d4SStefano Zampini 6113906d46d4SStefano Zampini PetscFunctionBegin; 6114906d46d4SStefano Zampini ierr = MatShellGetContext(A,&change_ctx);CHKERRQ(ierr); 6115906d46d4SStefano Zampini ierr = MatMult(change_ctx->global_change,x,change_ctx->work[0]);CHKERRQ(ierr); 6116906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->original_mat,change_ctx->work[0],change_ctx->work[1]);CHKERRQ(ierr); 6117906d46d4SStefano Zampini ierr = MatMultTranspose(change_ctx->global_change,change_ctx->work[1],y);CHKERRQ(ierr); 6118906d46d4SStefano Zampini PetscFunctionReturn(0); 6119906d46d4SStefano Zampini } 6120b96c3477SStefano Zampini 6121b96c3477SStefano Zampini #undef __FUNCT__ 6122b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCSetUpSubSchurs" 612308122e43SStefano Zampini PetscErrorCode PCBDDCSetUpSubSchurs(PC pc) 6124b96c3477SStefano Zampini { 6125a64f4aa4SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6126b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6127b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6128a64f4aa4SStefano Zampini Mat S_j; 6129b96c3477SStefano Zampini PetscInt *used_xadj,*used_adjncy; 6130b96c3477SStefano Zampini PetscBool free_used_adj; 6131b96c3477SStefano Zampini PetscErrorCode ierr; 6132b96c3477SStefano Zampini 6133b96c3477SStefano Zampini PetscFunctionBegin; 6134b96c3477SStefano Zampini /* decide the adjacency to be used for determining internal problems for local schur on subsets */ 6135b96c3477SStefano Zampini free_used_adj = PETSC_FALSE; 613608122e43SStefano Zampini if (pcbddc->sub_schurs_layers == -1) { 6137b96c3477SStefano Zampini used_xadj = NULL; 6138b96c3477SStefano Zampini used_adjncy = NULL; 6139b96c3477SStefano Zampini } else { 614008122e43SStefano Zampini if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) { 614108122e43SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 614208122e43SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 614308122e43SStefano Zampini } else if (pcbddc->computed_rowadj) { 6144b96c3477SStefano Zampini used_xadj = pcbddc->mat_graph->xadj; 6145b96c3477SStefano Zampini used_adjncy = pcbddc->mat_graph->adjncy; 6146b96c3477SStefano Zampini } else { 61472fffb893SStefano Zampini PetscBool flg_row=PETSC_FALSE; 6148b96c3477SStefano Zampini const PetscInt *xadj,*adjncy; 6149b96c3477SStefano Zampini PetscInt nvtxs; 6150b96c3477SStefano Zampini 61512fffb893SStefano Zampini ierr = MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 61522fffb893SStefano Zampini if (flg_row) { 6153b96c3477SStefano Zampini ierr = PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);CHKERRQ(ierr); 6154b96c3477SStefano Zampini ierr = PetscMemcpy(used_xadj,xadj,(nvtxs+1)*sizeof(*xadj));CHKERRQ(ierr); 6155b96c3477SStefano Zampini ierr = PetscMemcpy(used_adjncy,adjncy,(xadj[nvtxs])*sizeof(*adjncy));CHKERRQ(ierr); 6156b96c3477SStefano Zampini free_used_adj = PETSC_TRUE; 61572fffb893SStefano Zampini } else { 61582fffb893SStefano Zampini pcbddc->sub_schurs_layers = -1; 61592fffb893SStefano Zampini used_xadj = NULL; 61602fffb893SStefano Zampini used_adjncy = NULL; 61612fffb893SStefano Zampini } 61622fffb893SStefano Zampini ierr = MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);CHKERRQ(ierr); 6163b96c3477SStefano Zampini } 6164b96c3477SStefano Zampini } 6165d5574798SStefano Zampini 6166d5574798SStefano Zampini /* setup sub_schurs data */ 6167a64f4aa4SStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6168df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 6169df4d28bfSStefano Zampini /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */ 6170a64f4aa4SStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 617172b8c272SStefano 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); 6172a64f4aa4SStefano Zampini } else { 61736816873aSStefano Zampini PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis; 617404708bb6SStefano Zampini PetscBool isseqaij; 6175a3df083aSStefano Zampini PetscInt benign_n; 617672b8c272SStefano Zampini Mat change = NULL; 617772b8c272SStefano Zampini IS change_primal = NULL; 6178a3df083aSStefano Zampini 61795feab87aSStefano Zampini if (!pcbddc->use_vertices && reuse_solvers) { 61805feab87aSStefano Zampini PetscInt n_vertices; 61815feab87aSStefano Zampini 61825feab87aSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);CHKERRQ(ierr); 61832034aafcSStefano Zampini reuse_solvers = (PetscBool)!n_vertices; 61845feab87aSStefano Zampini } 618504708bb6SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 618604708bb6SStefano Zampini if (!isseqaij) { 618704708bb6SStefano Zampini Mat_IS* matis = (Mat_IS*)pc->pmat->data; 618804708bb6SStefano Zampini if (matis->A == pcbddc->local_mat) { 618904708bb6SStefano Zampini ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr); 619004708bb6SStefano Zampini ierr = MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 619104708bb6SStefano Zampini } else { 619204708bb6SStefano Zampini ierr = MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_REUSE_MATRIX,&pcbddc->local_mat);CHKERRQ(ierr); 619304708bb6SStefano Zampini } 619404708bb6SStefano Zampini } 6195a3df083aSStefano Zampini if (!pcbddc->benign_change_explicit) { 6196a3df083aSStefano Zampini benign_n = pcbddc->benign_n; 6197ca92afb2SStefano Zampini } else { 6198a3df083aSStefano Zampini benign_n = 0; 6199ca92afb2SStefano Zampini } 6200bf3a8328SStefano Zampini /* If the user defines additional constraints, we import them here. 6201bf3a8328SStefano 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 */ 620272b8c272SStefano Zampini if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) { 620388c03ad3SStefano Zampini PC_IS *pcisf; 620488c03ad3SStefano Zampini PC_BDDC *pcbddcf; 620588c03ad3SStefano Zampini PC pcf; 620688c03ad3SStefano Zampini 620788c03ad3SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 620888c03ad3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph"); 620988c03ad3SStefano Zampini } 621088c03ad3SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)pc),&pcf);CHKERRQ(ierr); 621188c03ad3SStefano Zampini ierr = PCSetOperators(pcf,pc->mat,pc->pmat);CHKERRQ(ierr); 621288c03ad3SStefano Zampini ierr = PCSetType(pcf,PCBDDC);CHKERRQ(ierr); 621388c03ad3SStefano Zampini /* hacks */ 621488c03ad3SStefano Zampini pcisf = (PC_IS*)pcf->data; 621572b8c272SStefano Zampini pcisf->is_B_local = pcis->is_B_local; 621672b8c272SStefano Zampini pcisf->vec1_N = pcis->vec1_N; 621772b8c272SStefano Zampini pcisf->BtoNmap = pcis->BtoNmap; 621872b8c272SStefano Zampini pcisf->n = pcis->n; 621972b8c272SStefano Zampini pcisf->n_B = pcis->n_B; 622088c03ad3SStefano Zampini pcbddcf = (PC_BDDC*)pcf->data; 622188c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->mat_graph);CHKERRQ(ierr); 622288c03ad3SStefano Zampini pcbddcf->mat_graph = pcbddc->mat_graph; 622388c03ad3SStefano Zampini pcbddcf->use_faces = PETSC_TRUE; 622488c03ad3SStefano Zampini pcbddcf->use_change_of_basis = PETSC_TRUE; 622588c03ad3SStefano Zampini pcbddcf->use_change_on_faces = PETSC_TRUE; 622672b8c272SStefano Zampini pcbddcf->use_qr_single = PETSC_TRUE; 622788c03ad3SStefano Zampini pcbddcf->fake_change = PETSC_TRUE; 622888c03ad3SStefano Zampini ierr = PCBDDCConstraintsSetUp(pcf);CHKERRQ(ierr); 622972b8c272SStefano Zampini /* store information on primal vertices and change of basis (in local numbering) */ 623072b8c272SStefano Zampini sub_schurs->change_with_qr = pcbddcf->use_qr_single; 623172b8c272SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);CHKERRQ(ierr); 623272b8c272SStefano Zampini change = pcbddcf->ConstraintMatrix; 623372b8c272SStefano Zampini pcbddcf->ConstraintMatrix = NULL; 623488c03ad3SStefano Zampini /* free unneeded memory allocated in PCBDDCConstraintsSetUp */ 623572b8c272SStefano Zampini ierr = PetscFree(pcbddcf->sub_schurs);CHKERRQ(ierr); 623688c03ad3SStefano Zampini ierr = MatNullSpaceDestroy(&pcbddcf->onearnullspace);CHKERRQ(ierr); 623788c03ad3SStefano Zampini ierr = PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);CHKERRQ(ierr); 623888c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->primal_indices_local_idxs);CHKERRQ(ierr); 623988c03ad3SStefano Zampini ierr = PetscFree(pcbddcf->onearnullvecs_state);CHKERRQ(ierr); 624088c03ad3SStefano Zampini ierr = PetscFree(pcf->data);CHKERRQ(ierr); 624188c03ad3SStefano Zampini pcf->ops->destroy = NULL; 624288c03ad3SStefano Zampini ierr = PCDestroy(&pcf);CHKERRQ(ierr); 624388c03ad3SStefano Zampini } 624472b8c272SStefano 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); 624572b8c272SStefano Zampini ierr = MatDestroy(&change);CHKERRQ(ierr); 624672b8c272SStefano Zampini ierr = ISDestroy(&change_primal);CHKERRQ(ierr); 6247ca92afb2SStefano Zampini } 6248d12d3064SStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6249b96c3477SStefano Zampini 6250b96c3477SStefano Zampini /* free adjacency */ 6251b96c3477SStefano Zampini if (free_used_adj) { 6252b96c3477SStefano Zampini ierr = PetscFree2(used_xadj,used_adjncy);CHKERRQ(ierr); 6253b96c3477SStefano Zampini } 6254b96c3477SStefano Zampini PetscFunctionReturn(0); 6255b96c3477SStefano Zampini } 6256b96c3477SStefano Zampini 6257b96c3477SStefano Zampini #undef __FUNCT__ 6258b96c3477SStefano Zampini #define __FUNCT__ "PCBDDCInitSubSchurs" 625908122e43SStefano Zampini PetscErrorCode PCBDDCInitSubSchurs(PC pc) 6260b96c3477SStefano Zampini { 6261b96c3477SStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6262b96c3477SStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6263b96c3477SStefano Zampini PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 6264b96c3477SStefano Zampini PCBDDCGraph graph; 6265b96c3477SStefano Zampini PetscErrorCode ierr; 6266b96c3477SStefano Zampini 6267b96c3477SStefano Zampini PetscFunctionBegin; 6268b96c3477SStefano Zampini /* attach interface graph for determining subsets */ 626908122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */ 62703301b35fSStefano Zampini IS verticesIS,verticescomm; 62713301b35fSStefano Zampini PetscInt vsize,*idxs; 6272b96c3477SStefano Zampini 6273b96c3477SStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);CHKERRQ(ierr); 62743301b35fSStefano Zampini ierr = ISGetSize(verticesIS,&vsize);CHKERRQ(ierr); 62753301b35fSStefano Zampini ierr = ISGetIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62763301b35fSStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);CHKERRQ(ierr); 62773301b35fSStefano Zampini ierr = ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);CHKERRQ(ierr); 62783301b35fSStefano Zampini ierr = ISDestroy(&verticesIS);CHKERRQ(ierr); 6279b96c3477SStefano Zampini ierr = PCBDDCGraphCreate(&graph);CHKERRQ(ierr); 62807fb0e2dbSStefano Zampini ierr = PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global);CHKERRQ(ierr); 62813301b35fSStefano Zampini ierr = PCBDDCGraphSetUp(graph,0,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);CHKERRQ(ierr); 62823301b35fSStefano Zampini ierr = ISDestroy(&verticescomm);CHKERRQ(ierr); 6283b96c3477SStefano Zampini ierr = PCBDDCGraphComputeConnectedComponents(graph);CHKERRQ(ierr); 6284b96c3477SStefano Zampini /* 6285b96c3477SStefano Zampini if (pcbddc->dbg_flag) { 6286b96c3477SStefano Zampini ierr = PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);CHKERRQ(ierr); 6287b96c3477SStefano Zampini } 6288b96c3477SStefano Zampini */ 6289b96c3477SStefano Zampini } else { 6290b96c3477SStefano Zampini graph = pcbddc->mat_graph; 6291b96c3477SStefano Zampini } 6292b96c3477SStefano Zampini 6293b96c3477SStefano Zampini /* sub_schurs init */ 6294a64f4aa4SStefano Zampini ierr = PCBDDCSubSchursInit(sub_schurs,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap);CHKERRQ(ierr); 6295a64f4aa4SStefano Zampini 6296b96c3477SStefano Zampini /* free graph struct */ 629708122e43SStefano Zampini if (pcbddc->sub_schurs_rebuild) { 6298b96c3477SStefano Zampini ierr = PCBDDCGraphDestroy(&graph);CHKERRQ(ierr); 6299b96c3477SStefano Zampini } 6300b96c3477SStefano Zampini PetscFunctionReturn(0); 6301b96c3477SStefano Zampini } 6302fa34dd3eSStefano Zampini 6303fa34dd3eSStefano Zampini #undef __FUNCT__ 6304fa34dd3eSStefano Zampini #define __FUNCT__ "PCBDDCCheckOperator" 6305fa34dd3eSStefano Zampini PetscErrorCode PCBDDCCheckOperator(PC pc) 6306fa34dd3eSStefano Zampini { 6307fa34dd3eSStefano Zampini PC_IS *pcis=(PC_IS*)pc->data; 6308fa34dd3eSStefano Zampini PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 6309fa34dd3eSStefano Zampini PetscErrorCode ierr; 6310fa34dd3eSStefano Zampini 6311fa34dd3eSStefano Zampini PetscFunctionBegin; 6312fa34dd3eSStefano Zampini if (pcbddc->n_vertices == pcbddc->local_primal_size) { 6313fa34dd3eSStefano Zampini IS zerodiag = NULL; 63144f1b2e48SStefano Zampini Mat S_j,B0_B=NULL; 6315fa34dd3eSStefano Zampini Vec dummy_vec=NULL,vec_check_B,vec_scale_P; 63164f1b2e48SStefano Zampini PetscScalar *p0_check,*array,*array2; 631775c01103SStefano Zampini PetscReal norm; 6318fa34dd3eSStefano Zampini PetscInt i; 6319fa34dd3eSStefano Zampini 6320fa34dd3eSStefano Zampini /* B0 and B0_B */ 6321fa34dd3eSStefano Zampini if (zerodiag) { 6322fa34dd3eSStefano Zampini IS dummy; 6323fa34dd3eSStefano Zampini 63244f1b2e48SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);CHKERRQ(ierr); 63254f1b2e48SStefano Zampini ierr = MatGetSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 6326fa34dd3eSStefano Zampini ierr = MatCreateVecs(B0_B,NULL,&dummy_vec);CHKERRQ(ierr); 6327fa34dd3eSStefano Zampini ierr = ISDestroy(&dummy);CHKERRQ(ierr); 6328fa34dd3eSStefano Zampini } 6329fa34dd3eSStefano Zampini /* I need a primal vector to scale primal nodes since BDDC sums contibutions */ 6330fa34dd3eSStefano Zampini ierr = VecDuplicate(pcbddc->vec1_P,&vec_scale_P);CHKERRQ(ierr); 6331fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->vec1_P,1.0);CHKERRQ(ierr); 6332fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6333fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6334fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6335fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6336fa34dd3eSStefano Zampini ierr = VecReciprocal(vec_scale_P);CHKERRQ(ierr); 6337fa34dd3eSStefano Zampini /* S_j */ 6338fa34dd3eSStefano Zampini ierr = MatCreateSchurComplement(pcis->A_II,pcis->A_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);CHKERRQ(ierr); 6339fa34dd3eSStefano Zampini ierr = MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);CHKERRQ(ierr); 6340fa34dd3eSStefano Zampini 6341fa34dd3eSStefano Zampini /* mimic vector in \widetilde{W}_\Gamma */ 6342fa34dd3eSStefano Zampini ierr = VecSetRandom(pcis->vec1_N,NULL);CHKERRQ(ierr); 6343fa34dd3eSStefano Zampini /* continuous in primal space */ 6344fa34dd3eSStefano Zampini ierr = VecSetRandom(pcbddc->coarse_vec,NULL);CHKERRQ(ierr); 6345fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6346fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6347fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 63484f1b2e48SStefano Zampini ierr = PetscCalloc1(pcbddc->benign_n,&p0_check);CHKERRQ(ierr); 63494f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i]; 6350fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6351fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6352fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6353fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6354fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6355fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6356fa34dd3eSStefano Zampini ierr = VecDuplicate(pcis->vec2_B,&vec_check_B);CHKERRQ(ierr); 6357fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec2_B,vec_check_B);CHKERRQ(ierr); 6358fa34dd3eSStefano Zampini 6359fa34dd3eSStefano Zampini /* assemble rhs for coarse problem */ 6360fa34dd3eSStefano Zampini /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */ 6361fa34dd3eSStefano Zampini /* local with Schur */ 6362fa34dd3eSStefano Zampini ierr = MatMult(S_j,pcis->vec2_B,pcis->vec1_B);CHKERRQ(ierr); 6363fa34dd3eSStefano Zampini if (zerodiag) { 6364fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 63654f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i]; 6366fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6367fa34dd3eSStefano Zampini ierr = MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);CHKERRQ(ierr); 6368fa34dd3eSStefano Zampini } 6369fa34dd3eSStefano Zampini /* sum on primal nodes the local contributions */ 6370fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6371fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6372fa34dd3eSStefano Zampini ierr = VecGetArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6373fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6374fa34dd3eSStefano Zampini for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]]; 6375fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array2);CHKERRQ(ierr); 6376fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcis->vec1_N,&array);CHKERRQ(ierr); 6377fa34dd3eSStefano Zampini ierr = VecSet(pcbddc->coarse_vec,0.);CHKERRQ(ierr); 6378fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6379fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6380fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6381fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 6382fa34dd3eSStefano Zampini ierr = VecGetArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6383fa34dd3eSStefano Zampini /* scale primal nodes (BDDC sums contibutions) */ 6384fa34dd3eSStefano Zampini ierr = VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);CHKERRQ(ierr); 6385fa34dd3eSStefano Zampini ierr = VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);CHKERRQ(ierr); 6386fa34dd3eSStefano Zampini ierr = VecRestoreArray(pcbddc->vec1_P,&array);CHKERRQ(ierr); 6387fa34dd3eSStefano Zampini ierr = VecAssemblyBegin(pcis->vec1_N);CHKERRQ(ierr); 6388fa34dd3eSStefano Zampini ierr = VecAssemblyEnd(pcis->vec1_N);CHKERRQ(ierr); 6389fa34dd3eSStefano Zampini ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6390fa34dd3eSStefano Zampini ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 6391fa34dd3eSStefano Zampini /* global: \widetilde{B0}_B w_\Gamma */ 6392fa34dd3eSStefano Zampini if (zerodiag) { 6393fa34dd3eSStefano Zampini ierr = MatMult(B0_B,pcis->vec2_B,dummy_vec);CHKERRQ(ierr); 6394fa34dd3eSStefano Zampini ierr = VecGetArray(dummy_vec,&array);CHKERRQ(ierr); 63954f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i]; 6396fa34dd3eSStefano Zampini ierr = VecRestoreArray(dummy_vec,&array);CHKERRQ(ierr); 6397fa34dd3eSStefano Zampini } 6398fa34dd3eSStefano Zampini /* BDDC */ 6399fa34dd3eSStefano Zampini ierr = VecSet(pcis->vec1_D,0.);CHKERRQ(ierr); 6400fa34dd3eSStefano Zampini ierr = PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);CHKERRQ(ierr); 6401fa34dd3eSStefano Zampini 6402fa34dd3eSStefano Zampini ierr = VecCopy(pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr); 6403fa34dd3eSStefano Zampini ierr = VecAXPY(pcis->vec1_B,-1.0,vec_check_B);CHKERRQ(ierr); 6404fa34dd3eSStefano Zampini ierr = VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);CHKERRQ(ierr); 6405fa34dd3eSStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm); 64064f1b2e48SStefano Zampini for (i=0;i<pcbddc->benign_n;i++) { 64074f1b2e48SStefano Zampini PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%d] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i])); 6408fa34dd3eSStefano Zampini } 64094f1b2e48SStefano Zampini ierr = PetscFree(p0_check);CHKERRQ(ierr); 6410fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_scale_P);CHKERRQ(ierr); 6411fa34dd3eSStefano Zampini ierr = VecDestroy(&vec_check_B);CHKERRQ(ierr); 6412fa34dd3eSStefano Zampini ierr = VecDestroy(&dummy_vec);CHKERRQ(ierr); 6413fa34dd3eSStefano Zampini ierr = MatDestroy(&S_j);CHKERRQ(ierr); 6414fa34dd3eSStefano Zampini ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 6415fa34dd3eSStefano Zampini } 6416fa34dd3eSStefano Zampini PetscFunctionReturn(0); 6417fa34dd3eSStefano Zampini } 6418